Meta Muse Voice Transcribe: Benchmarks, Pricing, API Limits, and What It Changes

Verdict: Muse Voice Transcribe is the most compelling new hosted speech API at launch for teams that value low cost, English accuracy, and native conversation events. In Kingy.ai’s bounded 30-minute public-audio test, Muse cut English word errors by 22.2% relative versus a local Whisper large-v3-turbo baseline. The same test also exposed the caveat buyers most need to see: a Hindi-English mixed-script transcript scored much worse than Whisper, while our diarization and endpointing proxies were not reliable enough for unattended high-stakes use.

Information, prices, and live API behavior checked on September 1, 2026. Kingy.ai paid for and ran the bounded test documented below; we did not receive review access, sponsorship, or compensation from Meta. Raw outputs, scripts, hashes, and limitations are preserved in the downloadable audit package.

Key takeaways

• Muse costs $0.18 per processed audio hour, or $3 per 1,000 minutes. Meta charges the same for streaming and file transcription, includes its zero-data-retention option at price parity, and bills down to whole processed seconds.¹
• In our 25-minute English aggregate, Muse scored 11.93% WER versus 15.34% for Whisper large-v3-turbo Q5_0—a 3.41-point absolute and 22.2% relative error reduction. Muse won every English slice in this small test.²
• On one five-minute Hindi-English tutorial, Muse scored 54.21% mixed-script WER-style versus Whisper’s 30.84% because Muse rendered many English software terms in Devanagari. Hindi + English language bias did not improve it. This is one script-sensitive clip, not a general Hindi ranking.³
• Artificial Analysis measured 3.0623% final streaming WER and 0.163 seconds from detected end of speech to final transcript on its English-weighted test. That was the best final WER in its September 1 snapshot—but not proof of universal accuracy.⁴⁵
• Meta’s launch chart reports 17.5% average diarization error rate across AMI-IHM, AMI-SDM, and VoxConverse. It is a provider-published comparison, not an independently reproduced result, and 17.5% is still too much speaker error for unattended high-stakes records.⁶
• Meta trained Muse on 70-plus languages but recommends 25 validated languages. Public launch material does not provide broad per-language, accent, dialect, or code-switch scores.⁷⁸
• The model learned transcription, diarization, and endpointing as token-generation tasks. In the present API, however, endpointing and diarization are separate session modes; buyers should not assume one connection returns both simultaneously.⁹
• Muse is a hosted Meta Model API service. Meta announced no downloadable Muse Voice Transcribe weights or self-hosted/on-premises route. That makes Deepgram or an open-weight stack more appropriate when local control is mandatory.¹⁰¹¹

Muse Voice Transcribe at a glance

Item	Launch/current fact	Evidence status
Release	September 1, 2026; Meta Superintelligence Labs	Meta announcement12
Model/API ID	muse-voice-transcribe-1.0	Meta developer docs13
Access	Hosted Meta Model API; also used in Meta AI for Mac and Muse Code	Provider-reported; no local weights announced1415
Protocols	WebSocket realtime plus multipart file transcription	Meta developer docs16
Price	$0.003/minute; $0.18/hour; $3/1,000 minutes	Meta pricing page17
Languages	70+ trained; 25 validated/recommended	Provider-reported; no broad public multilingual benchmark18
Validated languages	Arabic, Bengali, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Mandarin Chinese, Marathi, Polish, Portuguese, Spanish, Tagalog, Tamil, Telugu, Thai, Turkish, Vietnamese	Meta developer docs19
Streaming audio	Raw mono signed 16-bit little-endian PCM; 24 kHz preferred, 16 kHz accepted; realtime pacing required	Meta developer docs20
Current limits	Up to 60 minutes per realtime session; 10 minutes and 32 MB per file; eight concurrent streams and 1,000 starts/hour per tenant	Meta developer docs and rate limits2122
Conversation modes	Push-to-talk, endpointing, or diarization; selected per session	Meta developer docs23
Independent English result	3.0623% final WER; 0.163 s finalization latency after benchmark-detected end of speech	Artificial Analysis2425
Kingy.ai English test	11.93% WER vs 15.34% for local Whisper large-v3-turbo Q5_0; 25 public minutes	Independent, bounded test; not an AA reproduction26
Kingy.ai mixed-script test	54.21% Muse vs 30.84% Whisper WER-style; one 5-minute Hindi-English clip	Independent, script-sensitive result; not a broad language ranking27
Provider diarization result	17.5% average DER across three datasets	Meta-published; methodology incompletely disclosed28

What Muse Voice Transcribe is

Automatic speech recognition, or ASR, converts speech into text. A streaming ASR system does that while someone is still speaking instead of waiting for a completed recording. Diarization answers a different question—who spoke when?—while endpointing decides when a person has probably finished a turn.

Most voice products assemble those functions as a pipeline. A voice-activity detector notices sound; an endpoint detector decides the turn is over; an ASR model writes the words; a diarizer assigns speakers; timestamp and formatting stages clean the result. Every boundary creates timing, cost, and failure-handling work.

Muse is Meta’s attempt to teach one autoregressive multimodal model to express words and conversational structure in one token language. That is the strategic point. The launch is not merely “another transcription model with 0.3 points less WER.” It suggests that text, speaker turns, and speech boundaries can be native outputs of the same learned system.

The product boundary is narrower than the research story. Meta’s current API makes clients choose push-to-talk, endpointing, or diarization when opening a session. The diarization mode is explicitly not tuned for low-latency command interaction; Meta recommends endpointing for that job.²⁹ So Muse unifies the underlying model and price, but it does not yet give developers every conversation signal in one simultaneous API mode.

How the model works

Meta says Muse consumes audio in 80-millisecond chunks, or 12.5 steps per second. Each chunk becomes one continuous “soft token.” At every step, the model can emit transcript or task tokens, or request another audio chunk. An empty-audio token tells it the stream has ended and gives it a chance to finish.³⁰

That listen-or-write loop addresses a basic streaming dilemma. Commit too early and “recognize speech” can become “wreck a nice beach.” Wait too long and the interface feels sluggish. Meta calls its variable wait adaptive delay. It reports using reinforcement learning with a word-error reward and a delay reward combined multiplicatively, allowing the model to wait for ambiguous speech while emitting obvious words sooner.³¹

Special tokens extend the same sequence beyond text. Start-of-turn and speaker-label tokens support diarization; speech-onset and endpoint tokens support turn detection. This is appealing because the word choice, speaker boundary, and pause are inferred from shared context rather than reconciled after separate models disagree.

Important internals remain undisclosed. Meta has not published the parameter count, full training-data composition, training compute, native context size, per-language training mix, or a safety evaluation. “Part of the Muse Spark family” does not establish that Muse Voice Transcribe can reason over audio, identify people, detect emotion, or understand non-speech events. The public API specifically does not return confidence scores, word-level timestamps, sound-event labels, emotion, or formatted transcripts.³²

What is actually new—and what is packaging

Learned endpointing, streaming diarization, contextual biasing, multilingual ASR, and low-latency partials all existed before Muse. Several competitors already expose some combination. The novelty is their representation as generative events from one multimodal model at an unusually low hosted price, plus Meta’s ability to distribute the capability through its own desktop AI and coding products.

That could simplify a voice stack in three ways:

1. Fewer synchronization boundaries. A speaker event and transcript come from one stateful model rather than timestamps reconciled across services.
2. Less duplicated inference. One audio representation can inform text and turn structure.
3. Better semantic timing. Adaptive delay and endpointing can use linguistic context instead of only a fixed silence threshold.

The modular counterargument is strong. Separate components can be replaced, run in different regions, tuned for telephony versus meetings, and degraded independently. A unified hosted model concentrates correlated failures and vendor lock-in. Muse’s own mutually exclusive endpointing and diarization modes also mean many production teams will still need orchestration or an additional component.

Benchmark audit: a real lead, within a narrow test

Streaming English transcription

Artificial Analysis’ September 1 live table placed Muse first on final streaming WER. Its index uses roughly eight hours of English: about 250 minutes of proprietary held-out AA-AgentTalk voice-agent speech, 119 minutes of English VoxPopuli, and 115 minutes of Earnings22 calls. The index weights those components 50%, 25%, and 25%, respectively.³³

Streaming model/configuration	Final WER	First-partial WER	Final latency after detected endpoint	First-partial latency	AA normalized price / 1,000 min
Meta Muse Voice Transcribe	3.0623%	3.5747%	0.163 s	0.127 s	$3.00
Cartesia Ink-2, semantic endpoint	3.3612%	4.8878%	0.431 s	0.168 s	$4.00
ElevenLabs Scribe v2 Realtime	3.5946%	3.5928%	0.141 s	0.134 s	$6.50
Qwen3 ASR Flash Realtime	3.7339%	19.9445%	0.476 s	0.401 s	$5.40
OpenAI GPT Live Transcribe	3.9177%	6.3477%	0.812 s	0.256 s	$17.00
Grok Speech to Text Streaming	3.9329%	18.2751%	0.373 s	0.245 s	$3.33
Google Gemini 3.5 Transcribe Live	3.9983%	5.7743%	0.395 s	0.250 s	$9.00
AssemblyAI U3.5 Realtime Pro, minimum latency	4.0180%	4.0412%	0.191 s	0.185 s	$7.50

Snapshot: September 1, 2026. Values are live Artificial Analysis model/configuration results; prices are its normalized 1,000-minute estimates, not necessarily each vendor’s on-demand list rate.³⁴³⁵

WER is (substitutions + insertions + deletions) / reference words. A 3.1% WER is not literally “96.9% accurate”: one recognition mistake can affect multiple edits, and WER ignores whether an error changes meaning. Artificial Analysis normalizes outputs with the Whisper English normalizer plus additional equivalences, lowercases text, removes certain filler and bracketed tags, and standardizes punctuation, numbers, and spelling. That makes providers more comparable while hiding dimensions users notice—capitalization, verbatim punctuation, entity spelling, timestamp quality, and speaker attribution.³⁶

The measured lead over Cartesia’s semantic-endpoint configuration is about 0.30 absolute percentage points, or roughly an 8.9% relative reduction in word errors. At millions of words that can matter. But the leaderboard does not publish statistical error bars for this streaming table, and a single English corpus mixture cannot prove broad superiority. Endpoint configuration also changes the frontier: Cartesia’s externally forced endpoint result is much faster than its semantic configuration but has higher final WER.

The 0.163-second result is not total voice-agent latency. It begins at Artificial Analysis’ detected end of speech and ends when the final transcript arrives. A real response also includes microphone buffering, network transit, endpoint detection, reasoning, tools, safety checks, and speech synthesis. Muse is accurate and fast on this protocol; it does not make an entire agent respond in 163 milliseconds.

Speaker diarization

System/configuration	Average DER	Mode	Evidence owner
Muse Voice Transcribe	17.5%	Streaming	Meta launch chart
AssemblyAI U3.5 Pro	21.1%	Offline	Meta launch chart
ElevenLabs Scribe v2	24.6%	Offline	Meta launch chart
Deepgram Nova-3	25.4%	Offline	Meta launch chart
AssemblyAI U3.5 Pro	27.6%	Streaming	Meta launch chart
Deepgram Nova-3	28.6%	Streaming	Meta launch chart

Meta reports the simple average across AMI-IHM, AMI-SDM, and VoxConverse.³⁷ Meta did not publish per-dataset scores, scoring scripts, collar, overlap treatment, preprocessing, label-revision policy, network delay, or every competitor configuration. Treat the ranking as provider-published.

AMI is a 100-hour meeting corpus recorded with close-talk and distant microphones.³⁸ VoxConverse was built from varied “in the wild” video with overlapping speech, many speakers, and challenging background conditions.³⁹ DER normally combines missed speech, false speech, and speaker confusion after optimally mapping anonymous speaker labels. Results can move materially with the scoring collar, overlap rules, oracle versus estimated speaker counts, and channel choice.

Seventeen-and-a-half percent is first place in Meta’s chart, not a solved problem. A system can transcribe the right sentence and assign it to the wrong person. That is inconvenient in meeting notes and dangerous in clinical records, legal interviews, employee investigations, or an agent deciding who authorized an action. Muse’s A/B/C labels are session-local anonymous clusters, not verified identities; its docs warn that speaker events can lag the audio and do not by themselves guarantee a clean speaker boundary.⁴⁰

Kingy.ai independent test: Muse versus local Whisper

We ran a deliberately small, reproducible test rather than pretending a $0.17 experiment can settle the market. The frozen corpus contains 30 unique public minutes: 10 minutes of LibriSpeech test-clean, 15 minutes from the four-speaker AMI ES2004a Mix-Headset channel, and a five-minute MUCS Hindi-English LibreOffice tutorial.⁴¹⁴²⁴³ The AMI audio is mixed close-microphone audio—not far-field.

Muse transcribed all 30 minutes in batch mode. We separately ran its diarization mode on the 15 AMI minutes, its realtime endpointing mode on one five-minute AMI segment, and one extra Hindi + English language-bias pass. The zero-cost comparator was Whisper large-v3-turbo Q5_0 through whisper.cpp, CPU-only with Silero VAD 6.2.0. This is an available local baseline, not a like-for-like comparison with the strongest paid streaming competitor.

The result

Public-audio test slice	Duration	Meta Muse	Local Whisper	Winner
LibriSpeech test-clean	10 min	1.09% WER	1.87%	Muse
AMI meeting, minutes 0–5	5 min	15.17% WER	22.01%	Muse
AMI meeting, minutes 5–10	5 min	24.90% WER	29.12%	Muse
AMI meeting, minutes 10–15	5 min	17.09% WER	22.29%	Muse
English aggregate	25 min	11.93% WER	15.34%	Muse
MUCS Hindi-English	5 min	54.21% WER-style	30.84%	Whisper
Muse with Hindi + English bias	5 min extra	54.39% WER-style	—	No improvement

Across 3,840 English reference words, Muse made 458 edit errors versus Whisper’s 589. That is a 3.41 percentage-point absolute and 22.2% relative reduction in word errors. Muse won every English slice. The clean-read result is excellent; the meeting result is more useful because it shows both systems degrading on spontaneous multi-speaker speech.

The Hindi-English reversal needs careful interpretation. The reference preserves English technical terms such as “slide,” “font,” and “workspace” in Latin script. Whisper largely did too; Muse transliterated many into Devanagari. Standard token matching treats those script variants as errors even when they may sound similar. For products that need exact searchable software terms, that output convention is a real defect. It does not establish that Muse is generally worse at Hindi, multilingual speech, or semantic comprehension. The language-bias hint moved the score 0.19 points in the wrong direction—effectively no help in this clip.

Speaker and turn events: useful, not authoritative

Our custom AMI scorer measured 37.31% aggregate proxy DER and 45.98% macro JER across the 15 diarization minutes. It used 10 ms frames, no scoring collar, overlap included, and optimal anonymous-speaker mapping. Those settings are intentionally disclosed because diarization scores move with them. This is not comparable to Meta’s 17.5% provider chart, which averages different AMI channels plus VoxConverse under an incompletely published protocol.⁴⁴

On the five-minute realtime endpointing run, Muse produced 349 partial transcripts and 30 completed turns. The first valid partial arrived 573 ms after its relevant speech-start event, and partial churn was 3.49 edits per 100 prior tokens. A simple reference-activity matcher counted 9 false endpoint turns and 8 missed activity regions; matched end boundaries averaged 3.31 seconds from the annotated boundary. One early false turn occurred during initial silence. This one-call proxy is diagnostic, not a standard latency or endpoint benchmark.

Method, cost, and reproducibility

We normalized text with Unicode NFKC, lowercase, punctuation removal, and collapsed whitespace, retained fillers, then computed token edit distance. The Hindi-English result is labelled WER-style because mixed scripts make the unit unstable. We did not freeze an entity list before the run, so we refused to add a post-hoc entity score.

Successful Muse processing totaled 55 audio-minutes, or US$0.165 at list price. Meta’s account usage page subsequently posted 3.3k audio-transcription seconds and CA$0.22 spend, consistent with the run after currency conversion and rounding. The runner enforced a US$10 hard stop, reserved cost before every call, and never retried automatically. One pre-transcript request failed with HTTP 402 before billing was configured; Meta says such failures are not billed. We preserved it anyway. The conservative ledger including that failed reservation is US$0.195.⁴⁵⁴⁶

Whisper’s default Metal path crashed in preflight. A CPU run without VAD then hallucinated a repeated sentence through silence; we retained that output and changed the scored baseline to CPU + VAD. Because that was a post-hoc correction, it may favor Whisper. The full disclosure matters more than manufacturing symmetry.

The independent test report and benchmark audit package contain the frozen protocol, corpus manifest and hashes, references, scripts, model ledger, raw outputs, failed response, spend ledger, and metrics. Credentials, model binaries, and source audio are excluded.⁴⁷

Price and unit economics

Muse is billed on processed audio, rounded down to whole seconds. A 30.4-second request bills as 30 seconds. Requests that fail before producing a transcript and rate-limited requests are not billed. Streaming and file transcription cost the same; Meta lists no minimum commitment. It does not offer its training-eligible discount tier for voice at launch.⁴⁸

Streaming service	$/min	$/hour	1,000 hours	100,000 hours	Difference vs Muse	Caveat
Inworld STT 1	$0.00139	$0.0834	$83	$8,340	53.7% lower	AA normalized plan price; Inworld on-demand list rate is $0.15/hour49
Meta Muse	$0.0030	$0.18	$180	$18,000	—	Same price streaming/file; ZDR parity50
xAI Grok STT streaming	$0.00333	$0.20	$200	$20,000	11.1% higher	Official streaming rate; VAD can skip non-speech51
Cartesia Ink-2	$0.0040	$0.24	$240	$24,000	33.3% higher	AA normalized subscription price; plan credits/concurrency vary52
Qwen3 ASR Flash Realtime	$0.0054	$0.324	$324	$32,400	80.0% higher	Singapore international rate; regions differ53
ElevenLabs Scribe v2 Realtime	$0.0065	$0.39	$390	$39,000	116.7% higher	Official API list rate; realtime lacks diarization5455
Deepgram Flux	$0.0065	$0.39	$390	$39,000	116.7% higher	AA normalized; self-hosting available under enterprise terms56
AssemblyAI U3.5 RT Pro	$0.0075	$0.45	$450	$45,000	150.0% higher	AA normalized current configuration57
Gemini 3.5 Transcribe Live	$0.0090	$0.54	$540	$54,000	200.0% higher	Estimated blended token rate58
OpenAI GPT Live Transcribe	$0.0170	$1.02	$1,020	$102,000	466.7% higher	Official per-minute rate59

At Muse’s headline rate, 10 hours cost $1.80, 1,000 hours cost $180, and 100,000 hours cost $18,000. The arithmetic is simple; the bill is not the whole system. Transport, telephony, recording storage, analytics, an LLM, tools, retrieval, moderation, and TTS often dominate. Silence policy also matters: Meta bills processed audio, while xAI documents a voice-activity gate and other vendors structure credits or token charges differently.

There is no universal price/performance winner. Inworld is cheaper under Artificial Analysis’ normalized plan assumptions. xAI is close to Muse on price. Cartesia can deliver extremely fast externally forced endpoints. ElevenLabs supports more languages and word timestamps. Deepgram offers self-hosted deployment. Muse’s position is unusually balanced—lowest English final WER in this snapshot, competitive final latency, and near-bottom pricing—but buyers should choose the frontier that matches their audio and controls.

API and deployment reality

Meta exposes a WebSocket at wss://api.meta.ai/v1/asr/realtime and a multipart POST /v1/asr/transcribe endpoint. Realtime authentication happens in the WebSocket handshake; an Authorization header is ignored. Raw audio must be paced in realtime, and Meta allows only about five seconds of backlog. Sending a buffered hour as fast as possible is not supported.⁶⁰ Meta has promoted the API as a public preview, but the reviewed documentation does not publish a speech-specific region list or SLA.

Realtime sessions currently run for up to 60 minutes and cannot be resumed. A reconnect creates a new session, so the application must join transcript and speaker state. File transcription accepts mono 16-bit PCM WAV at 16 or 24 kHz, up to 32 MB and 10 minutes. Those production limits qualify the launch demo of a 1:00:52 conversation. Meta demonstrated hour-plus native handling; the current API does not accept a single file that long and does not promise an uninterrupted stream beyond 60 minutes.⁶¹⁶²

Partials are revisable. Endpointing events include speech start, partial transcript, speech end, and speech complete. Turns can overlap, so clients must key state by turn ID. The API does not provide a resume token, word timestamps, confidence scores, or speaker identity across sessions.⁶³

Capability	Muse	ElevenLabs Scribe v2 RT	Deepgram Flux	Inworld STT 1	Gemini 3.5 Live	OpenAI GPT Live Transcribe
Streaming	Yes	Yes	Yes	Yes	Yes	Yes
Native realtime diarization	Yes, separate session mode	No	Separate diarization feature; self-host constraints	Not clearly documented	Live model not listed with diarization	Not documented
Semantic endpointing	Yes, separate mode	Yes	Yes, integrated turn events	Automatic/manual turn taking	Live turn handling	Tunable latency
Language claim	25 validated; 70+ trained	90+	10-model multilingual Flux option	30+	Multilingual; no comparable validated count published	Multiple language hints
Code-switch evidence	Provider demos; no broad scores	Provider claim	Multilingual option	Not established here	Not established here	Multiple language hints
Long session	60 min/API session today	Provider-dependent	Reconnect patterns documented	Provider-dependent	Context/session rules apply	Provider-dependent
Word timestamps	No	Yes	Yes	Not established here	Batch sibling supports them	Not listed
Self-host/open weights	No announced route	No	Self-hosted enterprise path	Enterprise on-prem option	No	No
Retention/control	Meta names ZDR at same price; scope must be contracted	Enterprise controls vary	Self-host and enterprise controls	ZDR; HIPAA/BAA add-ons on higher plans	Paid API says inputs not used to improve products	API data controls apply

Capability labels are based on each provider’s public pages as checked September 1, 2026.^{64656667686970} “ZDR” is not the same as on-device processing, data residency, HIPAA eligibility, or permission to record.

Meta says audio in the launch-page demo is processed to generate a transcript and not stored. That sentence applies to the demo. The pricing page says a ZDR option exists for the Model API at standard price, but accessible public material does not define enough scope here to infer exact retention windows, abuse-monitoring exceptions, regions, certifications, or a service-level agreement. Buyers need those terms in writing.⁷¹⁷²

Can you run Muse on Vercel or Runpod?

You can run the integration there; you cannot run the Muse model there. Muse remains a Meta-hosted API. A Vercel or Runpod service would hold your application logic: issue short-lived client credentials where supported, proxy batch uploads, normalize events, stitch sessions, enforce budgets, store consented outputs, or run evaluation jobs. A GPU does not accelerate Muse inference because the audio still goes to Meta.

Vercel is a clean fit for a web UI and short batch-control path. The realtime path needs more care. Vercel’s current Functions limits still say Functions do not act as WebSocket servers, while Vercel Services documents WebSocket backends on Fluid compute and the June 2026 Functions update allows up to 30-minute Pro/Enterprise executions in beta.⁷³⁷⁴⁷⁵ Muse permits a 60-minute realtime session. Therefore a conventional Vercel Function cannot safely represent the full native session: use direct browser-to-Meta audio where security and the API design permit it, cap/reconnect shorter sessions, use Vercel Services, or put a persistent realtime broker elsewhere.

Runpod can host that persistent broker or the entire benchmark harness. Serverless endpoint execution time is configurable from five seconds to seven days, although the documented default is 600 seconds; Pods run persistent TCP/HTTP services and bill by the minute.⁷⁶⁷⁷ Use an inexpensive CPU worker or Pod for Muse orchestration. Choose a GPU only if the same deployment also runs Whisper or another local model. Persist raw outputs and ledgers on a network volume or external object store rather than ephemeral container disk.⁷⁸

A sensible production split is: Vercel for the product UI and ordinary API routes; a persistent regional WebSocket worker for Muse realtime; object storage for consented audio/outputs; and a queue for batch transcription and retries. Keep the Meta key server-side, never put a long-lived credential in browser JavaScript, and enforce cost and duration limits independently of Meta’s billing alert.

Where Muse could matter now

Use case	User value	Why Muse may fit	Critical metric	Integration pattern	Main failure mode	Required oversight/compliance	Readiness now
Voice agents/support	Faster natural turns	Low WER, endpoint events, low price	False endpoints; p95 response time	Endpoint mode → LLM/tools → TTS	Acts before user finishes	Confirm consequential actions; audit tools	Pilot
Contact-center assist	Live notes and suggestions	Cheap continuous streaming	Entity accuracy; disconnects	Stream audio; agent UI; CRM writeback	Wrong account/action	Consent, redaction, human approval	Pilot
Meeting notes	Searchable multi-speaker record	Native diarization	DER, overlap, label stability	Diarization mode → summarizer	Quote assigned to wrong speaker	Visible recording notice; review	Pilot with review
Dictation/accessibility	System-wide text input	Meta already ships Mac dictation	Correction rate; latency	Push-to-talk into focused field	Sensitive text sent to cloud	Clear cloud disclosure; undo	Ready for low-risk use
Live captions	Affordable low-delay text	Strong English streaming result	Stable partials; omissions	Endpoint/partial stream → caption UI	Partial churn confuses audience	Label automated captions	Pilot
Code-switched service	One conversation across languages	25 validated languages; code-switch demos	Per-pair WER/CER	Language bias + glossary	Names/transliteration fail	Native-speaker QA	Bake-off first
Voice coding/desktop agents	Hands-free commands	Keyword bias; low latency	Command intent; endpoint error	Endpoint mode → permissioned action	Spoken prompt injection/action	Read-back and permission gates	Pilot
AI glasses/wearables	Ambient context and recall	Low cost; long-session research	Battery/network; consent; DER	Rolling sessions → local policy → memory	Covert recording/misattribution	Prominent capture signals; bystander policy	Research only
Healthcare documentation	Less clerical burden	Diarization and terminology bias	Clinical entity and speaker error	Capture → structured draft → clinician sign-off	Medication/negation error	Consent, BAA, residency, clinician review	Wait / controlled eval
Legal interviews	Searchable record	Long-form and speakers	Verbatim fidelity; chain of custody	Redundant recording + draft transcript	Incorrect attribution becomes evidence	Jurisdictional consent; certified review	Not authoritative
Education/lectures	Notes and captions	Low cost and multilingual coverage	Names, formulas, accessibility delay	Stream → captions/notes	Technical terms omitted	Student consent; accessible fallback	Pilot
Field service/logistics	Hands-free notes	Streaming plus keyword bias	Noise WER; offline recovery	Headset → glossary → work order draft	Noise changes part numbers	Read-back; no safety actuation	Pilot after noise test
Gaming/multi-user spaces	Moderation, commands, captions	Many-speaker model claim	Overlap and speaker tracking	Diarization stream → moderation/UI	Speaker labels drift	Appeals; privacy and youth safeguards	Experimental
Journalism/podcasts	Faster rough transcripts	Low cost; meeting diarization	Quotes, names, timestamps	Diarization → editor → published copy	Fabricated or misassigned quote	Listen-back for every quote	Ready as a draft tool
Sales/CRM automation	Searchable calls and follow-up	Cheap streaming, glossary hints	Numbers, names, commitments	Assist UI → reviewed CRM update	False commitment logged	Recording consent; rep approval	Pilot
Live translation	Lower-latency source text	Multilingual/code-switch promise	Source WER plus translation quality	Muse → translation → captions/TTS	Two systems compound errors	Human fallback for critical content	Bake-off first
Robotics/in-room assistants	Better turn detection	Semantic endpoint mode	Barge-in; false activation	Endpoint → constrained intent layer	Wrong person triggers device	Wake word, identity, confirmation	Experimental

The best immediate fits are reversible, cloud-compatible workflows: captions, draft notes, dictation, agent assist, and voice interfaces where a mistaken transcript can be corrected. The weakest are workflows that treat a partial transcript or anonymous speaker label as authority.

What Muse means for AI products

Turn-taking is becoming a model capability

Silence timers make agents either interrupt or wait awkwardly. Semantic endpointing reframes the decision as: given the words, prosody, and context so far, is this thought complete? Adaptive delay applies the same principle within transcription. That moves conversational timing from hand-tuned glue into learned behavior—useful, but deserving its own false-endpoint and barge-in tests.

Audio perception is moving into multimodal model families

Muse is positioned inside Muse Spark rather than as a classic standalone recognizer. The architectural direction is plausible: multimodal foundation models will increasingly emit text, turns, and structured audio events from shared state. It is a bridge toward richer audio understanding, not proof that Meta has shipped intent, emotion, environmental-event recognition, or direct audio reasoning.

Streaming transcription is becoming commodity infrastructure

At $0.18/hour, ASR is a rounding error for many paid software products. Specialist vendors cannot rely on transcription alone. Their defensible value shifts to on-prem deployment, domain adaptation, telephony, enterprise support, compliance, analytics, timestamp fidelity, and operational reliability. Meta’s cross-product distribution—from Model API to desktop dictation and coding—adds pressure. Any claim that the rate is subsidized or designed to seed an ecosystem is inference; Meta has not said so.

Multilingual conversation is the prize—and the evidence gap

Within-sentence code-switching matters because real households, workplaces, and markets do not observe neat language boundaries. A model that keeps the entity and syntax intact across Mandarin and English, Hindi and English, or Spanish and English can remove language-selection UI. Yet a handful of demos cannot establish robustness. Teams need WER or character error rate by language, accent, dialect, code-switch pair, named entity, number, and transliteration convention.

Ambient AI raises a consent problem before a model problem

Low cost, long-context research, and many-speaker diarization point toward meeting companions and glasses that continuously interpret a room. Transcription is not memory: durable storage, identity across sessions, summarization, deletion, access control, and bystander consent are separate systems. Session-local “Speaker B” is not a person. Products that turn ambient speech into searchable memory need a social contract as explicit as their privacy architecture.

Developers must treat transcripts as revisable, untrusted events

Partials change. Turns overlap. Speaker events lag. Reconnects split state. Spoken text can contain prompt injection, including instructions aimed at a downstream agent rather than a human. False endpointing can trigger a tool too early; diarization error can assign an authorization to the wrong person; keyword bias lists can expose sensitive project names. Production systems need a state machine, provenance, permission checks, and explicit confirmation before side effects—not a text box wired directly to an agent.

Muse does not solve the rest of voice AI

Good transcription does not guarantee good reasoning, factual answers, expressive TTS, full-duplex interruption handling, reliable tools, memory, privacy, or user trust. Muse improves the “ears” and some timing signals. The brain, voice, hands, safeguards, and social permission remain product responsibilities.

Risks, limitations, and unanswered questions

• The tail matters. Aggregate WER can conceal catastrophic errors in names, negation, dosage, addresses, serial numbers, or currencies.
• Speaker attribution remains noisy. Meta’s own average DER is 17.5%, and the public protocol is incomplete.
• The multilingual claim is under-evaluated. There are no public scores for all 25 validated languages or systematic code-switch pairs.
• The API lacks common production signals. No confidence scores, word timestamps, formatted transcript, or cross-session speaker identity.
• Long context is qualified by service limits. Today’s 60-minute stream and 10-minute file cap require application stitching.
• Hosted-only access creates concentration risk. No public weights, on-prem path, or local fallback was announced.
• Enterprise details are incomplete. Meta has not publicly disclosed speech-specific regions, residency, certifications, uptime SLA, support terms, or the full operational scope of ZDR in the pages reviewed.
• Unified failures may correlate. If one model gets the boundary wrong, it may also get the nearby word and speaker wrong.

Muse versus the alternatives

Choose on the constraint that can kill your product, not on one leaderboard row.

• Muse is a strong candidate when low-cost hosted streaming, excellent English benchmark WER, anonymous speaker turns, or semantic endpoints matter—and the 25-language list covers the product.
• Deepgram deserves attention when self-hosting, telephony maturity, word timestamps, or independent component choice matters.⁷⁹
• ElevenLabs is attractive when 90-plus language coverage and word timestamps matter more than native realtime diarization.⁸⁰
• Cartesia belongs in a latency bake-off when externally controlled endpoints and very fast finalization are central.⁸¹
• Inworld is the cost challenger under subscription-normalized economics and offers enterprise on-prem and privacy add-ons.⁸²
• Gemini or a multimodal live model may be better when transcription is one feature inside a persistent audio-reasoning session rather than a standalone ASR service.⁸³
• An open-weight Whisper-family or other local stack remains preferable when offline processing, full data custody, or custom inference is non-negotiable, accepting more engineering and potentially weaker realtime behavior.

A serious evaluation plan

Build a consented test set from the product’s actual world: close-talk and far-field microphones; mobile and telephony codecs; reverberation, machinery, music, cross-talk, interruptions, and overlap; every supported accent, dialect, language, and code-switch pair; dates, money, addresses, alphanumerics, names, jargon, short commands, monologues, meetings, and stitched hour-plus sessions. Do not use customer or personal audio without authorization.

Measure more than WER:

• WER/CER, named-entity error, number/date/currency accuracy, omission, hallucination, punctuation, casing, timestamp error, and partial-revision rate.
• Time to first partial and stable final; p50/p95/p99; setup; regional network impact; and total agent response latency.
• DER/JER, speaker confusion, missed and false speech, overlap, label stability, and speaker-change delay.
• False and missed endpoints, premature actions, barge-in, backchannels, and long pauses.
• Processed-audio cost plus silence, retries, channels, storage, LLM, tools, TTS, and engineering.
• Disconnect recovery, rate limits, service availability, support, and auditability.

Send identical audio with equivalent preprocessing and region where possible. Freeze model IDs and dates. Run repeated trials and report variance. Preserve raw provider output before post-processing. Blind human reviewers to provider identity. Publish the corpus, configuration, scripts, and limitations when licensing permits.

For safety, never let a partial transcript, endpoint token, or inferred speaker identity alone authorize a consequential action. Apply confirmation, permissions, audit logs, and human review proportional to harm.

Who should use Muse now—and who should wait

Test now: voice-agent teams, contact-center platforms, meeting tools, captioning products, accessibility software, desktop dictation, and voice coding systems that can use a hosted API and keep a human in the loop. At $3 per 1,000 minutes, a representative bake-off is cheap.

Run a serious multi-vendor bake-off: products centered on accents, far-field audio, overlap, code-switching, domain entities, long sessions, or stable speaker attribution. The public benchmarks do not resolve those cases.

Wait or choose another path: teams requiring on-prem/local inference, downloadable weights, word-level timestamps, confidence scores, strict residency, a documented enterprise SLA/certification, or legally authoritative speaker records.

Muse earns a place on the shortlist. It has not earned blind migration.

The next 12–24 months

1. High confidence: Streaming ASR list prices cluster below $0.25/hour for high-volume tiers as vendors defend differentiation elsewhere.
2. High confidence: More APIs expose semantic endpoint and turn events alongside text; silence-only timers become a fallback.
3. Medium confidence: Meta adds a mode or event schema that combines endpointing and diarization in one session, closing the gap between model story and API surface.
4. Medium confidence: Independent leaderboards add multilingual and code-switch tracks because English WER stops separating top vendors.
5. Medium confidence: Enterprise buyers demand diarization-specific quality reports, revision semantics, and audit logs instead of accepting one DER average.
6. Medium confidence: At least one major voice vendor bundles ASR near cost and monetizes orchestration, TTS, analytics, or compliance.
7. Low confidence: Meta releases a smaller local Muse Voice model. Its open-weight Muse Glimmer shows that Meta can ship local models, but Voice Transcribe has no announced local path.⁸⁴

FAQ

How much does Muse Voice Transcribe cost?

$0.003 per processed minute, $0.18 per audio hour, or $3 per 1,000 minutes. Streaming and file transcription cost the same; Meta says ZDR is price-parity and free-tier credits apply.⁸⁵

Is Muse the most accurate speech model?

It ranked first for final WER in Artificial Analysis’ September 1 English streaming snapshot. That does not prove it is best across languages, accents, noisy rooms, domain vocabulary, formatting, or speaker attribution.⁸⁶⁸⁷

Is 3.1% WER the same as 96.9% accuracy?

No. WER counts substitutions, insertions, and deletions relative to reference words. It is a useful error rate, not a universal accuracy percentage or a measure of meaning preservation.

How fast is it?

Artificial Analysis measured 0.127 seconds to the first post-endpoint partial and 0.163 seconds to final text. Those clocks start after its detected end of speech, not when the user starts talking, and exclude downstream reasoning/TTS.⁸⁸⁸⁹

How many languages does Muse support?

Meta says it trained on 70-plus and recommends 25 validated languages. Do not treat those statements as evidence of equal quality across 70 languages.⁹⁰⁹¹

Does it support speaker diarization?

Yes, using anonymous session-local speaker labels. Meta reports 17.5% average DER on three datasets. The API’s diarization mode is separate from its low-latency endpointing mode.⁹²⁹³

What is endpointing?

Endpointing estimates that a speaker has finished a turn so an agent can respond. It is more consequential than a speed number: a false endpoint can interrupt a user or trigger a tool prematurely.

Are Muse’s weights open? Can it be self-hosted?

No downloadable Voice Transcribe weights or self-host/on-prem route were announced. It is currently a hosted Meta Model API.⁹⁴⁹⁵

Does Meta retain the audio?

Meta says its launch web demo processes audio for transcripts and does not store it, and its pricing page lists a ZDR API option. Those are different scopes. Buyers should verify retention, logging exceptions, residency, subprocessors, and contract terms for their account.⁹⁶⁹⁷

Can it transcribe an hour-long recording?

Meta demonstrated a 1:00:52 conversation and says the model natively handles hour-plus audio. The current realtime API caps a session at 60 minutes, and the file endpoint caps a file at 10 minutes, so applications must segment or reconnect and stitch.⁹⁸⁹⁹

What are the best early use cases?

Reversible, human-reviewable tasks: live captions, dictation, draft meeting notes, contact-center assist, voice interfaces, and rough media transcripts. Healthcare, legal, employment, financial, and safety actions require domain evaluation and human approval.

What did Kingy.ai’s independent test find?

Muse scored 11.93% WER versus 15.34% for a local Whisper large-v3-turbo Q5_0 baseline across 25 English minutes, a 22.2% relative error reduction. Whisper won one five-minute Hindi-English mixed-script clip, 30.84% to Muse’s 54.21% WER-style. The test is transparent but too small for a universal ranking.¹⁰⁰

Can Muse Voice Transcribe run on Vercel or Runpod?

The integration can; the Muse model cannot. Muse inference stays on Meta’s hosted API. Vercel can serve the UI and short request paths, while long realtime WebSocket sessions need Vercel Services, direct client connectivity where safe, reconnects, or a persistent broker. Runpod can host a worker or benchmark harness, but a GPU helps only if you also run a local model such as Whisper.¹⁰¹¹⁰²

Claim versus evidence

Claim	Evidence	Evidence owner	What it proves	What it does not prove	Confidence
Muse beat local Whisper on English	11.93% vs 15.34% WER over 25 public minutes	Kingy.ai	Better result under the frozen small protocol	Superiority to every competitor or domain	High, very narrow scope
Muse struggled with exact mixed-script fidelity	54.21% vs 30.84% WER-style on one MUCS clip	Kingy.ai	Script/output mismatch in this clip	General Hindi or code-switch quality	Medium, single clip
Muse diarization needs review	37.31% proxy DER over 15 AMI Mix-Headset minutes	Kingy.ai	High error under our disclosed scorer	Reproduction of Meta’s 17.5% chart	Medium, protocol-sensitive
Muse leads streaming WER	3.0623% AA-WER	Artificial Analysis	Best result in current English protocol	Universal or multilingual superiority	High, narrow scope
Muse is low cost	$0.18/hour official price	Meta	Launch list rate and billing rules	Total system cost or future pricing	High
Muse has strong diarization	17.5% average DER	Meta	Provider-observed first place on three datasets	Independent reproduction or safe attribution	Medium
One model handles ASR, diarization, endpoints	Architecture and task tokens	Meta	Shared generative model design	All events simultaneously in current API	High with API caveat
It supports code-switching	Launch demos and provider claim	Meta	Demonstrated capability in selected examples	Broad pair/accent robustness	Medium-low
It handles 70+ languages	Training statement	Meta	Languages appeared in training	Validation or equal accuracy	Medium
It handles hour-plus audio	1:00:52 demo and claim	Meta	Demonstrated research/product behavior	Current one-file or >60-minute API session	Medium
It is private	Demo no-store notice; ZDR option	Meta	Two scoped data-control statements	Universal retention, residency, compliance	Low until terms verified

Disclosure and sources

Kingy.ai purchased ordinary pay-as-you-go access and ran the bounded independent test on September 1, 2026. Meta did not provide review access, credits, sponsorship, editorial input, or compensation. We used only public CC-licensed audio and one zero-cost local comparator. Successful Muse calls processed 55 audio-minutes: US$0.165 by list-price calculation, posted as CA$0.22 on Meta’s usage page. A local hard cap prevented the paid ledger from exceeding US$10. Raw outputs and scripts were preserved before interpretation.

We did not reproduce Meta’s diarization chart or Artificial Analysis’ streaming benchmark, test customer/private audio, run a second paid provider, or claim broader hands-on product experience. Provider, independent-lab, and Kingy.ai results are labelled separately throughout. This page is maintained as a living guide; substantive changes will be logged below.

Update log

• September 1, 2026 — independent-test update: Added Kingy.ai’s 30-minute public-audio test, English and mixed-script transcript results, diarization and endpointing proxies, cost ledger, Vercel/Runpod deployment guidance, downloadable raw-output audit package, and revised disclosure.
• September 1, 2026: Initial publication. Launch-day Meta documentation, pricing, API limits, Artificial Analysis leaderboard/methodology, dataset sources, and competitor primary pages checked.

---

1. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

2. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

3. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

4. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

5. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

6. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

7. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

8. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

9. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

10. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

11. The New Stack, “Meta just beat OpenAI and Google at real-time transcription”, reporting a Meta spokesperson said weights would not be released.↩︎

12. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

13. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

14. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

15. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

16. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

17. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

18. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

19. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

20. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

21. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

22. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

23. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

24. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

25. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

26. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

27. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

28. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

29. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

30. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

31. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

32. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

33. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

34. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

35. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

36. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

37. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

38. AMI Meeting Corpus.↩︎

39. J. S. Chung et al., “Spot the conversation: speaker diarisation in the wild”, Interspeech 2020.↩︎

40. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

41. LibriSpeech ASR corpus, OpenSLR 12.↩︎

42. AMI Meeting Corpus.↩︎

43. MUCS 2021 Hindi-English speech corpus, OpenSLR 104.↩︎

44. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

45. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

46. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

47. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

48. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

49. Inworld pricing and Inworld Speech-to-Text, checked September 1, 2026.↩︎

50. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

51. xAI API pricing and Voice API overview, checked September 1, 2026.↩︎

52. Cartesia pricing, checked September 1, 2026.↩︎

53. Alibaba Cloud Model Studio pricing, Singapore international qwen3-asr-flash-realtime rate, checked September 1, 2026.↩︎

54. ElevenLabs API pricing, checked September 1, 2026.↩︎

55. ElevenLabs Realtime Speech-to-Text, checked September 1, 2026.↩︎

56. Deepgram Flux quickstart, checked September 1, 2026.↩︎

57. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

58. Google Gemini Developer API pricing, checked September 1, 2026.↩︎

59. OpenAI GPT Live Transcribe model page, checked September 1, 2026.↩︎

60. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

61. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

62. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

63. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

64. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

65. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

66. ElevenLabs Realtime Speech-to-Text, checked September 1, 2026.↩︎

67. Deepgram Flux quickstart, checked September 1, 2026.↩︎

68. Inworld pricing and Inworld Speech-to-Text, checked September 1, 2026.↩︎

69. Google Gemini Developer API pricing, checked September 1, 2026.↩︎

70. OpenAI GPT Live Transcribe model page, checked September 1, 2026.↩︎

71. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

72. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

73. Vercel, “Limits”, checked September 1, 2026.↩︎

74. Vercel, “Vercel Services, servers, and WebSockets with Fluid compute”, checked September 1, 2026.↩︎

75. Vercel, “Vercel Functions can now run up to 30 minutes,” June 15, 2026.↩︎

76. Runpod, “Endpoint configurations”, checked September 1, 2026.↩︎

77. Runpod, “Pods overview”, checked September 1, 2026.↩︎

78. Runpod, “Storage types”, checked September 1, 2026.↩︎

79. Deepgram Flux self-hosted, checked September 1, 2026.↩︎

80. ElevenLabs Realtime Speech-to-Text, checked September 1, 2026.↩︎

81. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

82. Inworld pricing and Inworld Speech-to-Text, checked September 1, 2026.↩︎

83. Google Gemini Live API overview, checked September 1, 2026.↩︎

84. Meta, “Introducing Muse Glimmer”, August 10, 2026.↩︎

85. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

86. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

87. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

88. Artificial Analysis, Streaming Speech-to-Text Leaderboard, snapshot September 1, 2026.↩︎

89. Artificial Analysis, Speech-to-Text Methodology v2.2, May 2026.↩︎

90. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

91. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

92. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

93. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

94. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

95. The New Stack, “Meta just beat OpenAI and Google at real-time transcription”, reporting a Meta spokesperson said weights would not be released.↩︎

96. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

97. Meta Model API, “Pricing and rate limits”, checked September 1, 2026.↩︎

98. Meta, “Introducing Muse Voice Transcribe,” September 1, 2026.↩︎

99. Meta Model API, “Speech to text” documentation, checked September 1, 2026.↩︎

100. Kingy.ai, “Independent test: Meta Muse Voice Transcribe 1.0”, September 1, 2026; download the audit package.↩︎

101. Vercel, “Vercel Services, servers, and WebSockets with Fluid compute”, checked September 1, 2026.↩︎

102. Runpod, “Pods overview”, checked September 1, 2026.↩︎