Meta costs Muse Voice Transcribe at $0.18 an hour, with real-time diarization for 20+ audio system: a steal for enterprises?
Meta is getting into the more and more aggressive real-time speech-to-text market with Muse Voice Transcribe, a brand new audio notion mannequin that mixes streaming transcription, endpoint detection and speaker diarization for greater than 20 audio system — at a public API worth of simply $0.18 per hour of processed audio.
Developed by Meta Superintelligence Labs, Muse is designed to course of speech whereas it occurs somewhat than ready for a recording to complete. Meta’s launch post for Muse Voice Transcribe says the mannequin helps lengthy audio exceeding an hour, seamless multilingual code-switching, language and key phrase biasing, and diarization with no separate post-processing pipeline. The mannequin was skilled throughout greater than 70 languages, with 25 extensively validated for the preliminary launch.
The 20-plus-speaker determine is substantial, however it’s not a planet document. A evaluate of present vendor documentation turns up methods with larger printed ceilings. Speechmatics' real-time transcription service says it could possibly determine 50 audio system by default and as much as 100 when the restrict is elevated, whereas Amazon Transcribe's diarization documentation specifies a most of 30 distinctive audio system, together with for streaming transcription. (Speechmatics)
Muse nonetheless lands towards the excessive finish of the market, and Meta's broader proposition is arguably extra essential than the uncooked most: high-capacity real-time diarization mixed with low-latency transcription, endpointing, multilingual code-switching and aggressive API pricing in the identical mannequin.
For enterprise builders constructing assembly methods, name analytics, stay assistants or ambient AI, that mixture might matter greater than who holds the speaker-count document.
Diarization is changing into a part of the core voice stack
Traditional speech recognition solutions a comparatively easy query: What was mentioned? Diarization provides one other: Who mentioned it?
That distinction turns into important as transcripts feed downstream AI methods. A gathering assistant can accurately transcribe each sentence and nonetheless create an unreliable company document if it attributes an approval, dedication or objection to the flawed participant. The identical situation impacts customer-service analytics, compliance workflows and AI brokers working in rooms the place a number of individuals can converse.
Muse incorporates speaker attribution straight into its autoregressive multimodal structure. Meta says audio arrives in 80-millisecond chunks, or 12.5 chunks per second, with every remodeled right into a gentle token. At every step, the mannequin decides whether or not to devour extra audio or emit textual content. Meta calls this mechanism adaptive delay: somewhat than making use of one latency funds to each phrase, Muse can wait longer when speech is ambiguous and commit earlier when it has sufficient context. Meta says reinforcement studying combines word-error-rate and delay rewards to coach that conduct. Meta's technical explanation of Muse particulars the structure. (Meta AI Research)
Speaker attribution and endpointing then develop into a part of the identical token sequence. A <|start_of_turn|> token marks a possible new speaker flip, tokens corresponding to <|speaker_A|> determine the speaker, and separate onset and endpoint tokens determine speech boundaries. Meta says it trains ASR, diarization and endpointing collectively somewhat than working speaker clustering as an unrelated downstream course of.
Meta's Model API speech-to-text documentation additionally exposes diarization as a first-class working mode alongside push-to-talk and endpointing. Speaker labels corresponding to A and B are scoped to a session somewhat than verified identities, and the API offers turn-level somewhat than word-level timestamps.
20+ audio system is excessive, however Speechmatics goes significantly larger
Speaker-count comparisons require care as a result of distributors implement diarization in a different way and don’t all publish a most.
Speechmatics presently makes the strongest express real-time capability declare discovered on this evaluate. Its real-time STT documentation says speaker diarization is out there stay, whereas its real-time FAQ says the system helps 50 audio system by default and will be elevated to 100.
AWS likewise exceeds Meta's acknowledged determine: Amazon Transcribe can differentiate a most of 30 distinctive audio system, and AWS offers express directions for speaker partitioning in a streaming transcription.
Soniox helps diarization in each real-time and asynchronous processing, however paperwork a most of 15 audio system per session. AssemblyAI's streaming diarization system lets builders set max_speakers between one and 10. Both corporations warning that stay speaker attribution is harder as a result of streaming methods should make selections with much less future audio context than offline fashions.
xAI's current Speech-to-Text API additionally helps speaker diarization in streaming mode, however its documentation reviewed for this story doesn’t publish a most diarized-speaker depend, so a direct ceiling comparability with Muse will not be attainable. (X.ai Docs)
That means it could be inaccurate to explain Muse's 20-plus functionality as a brand new world document. The highest explicitly documented real-time quantity recognized on this survey is Speechmatics' configurable 100-speaker ceiling.
Meta additionally doesn’t exhibit 20-plus simultaneous contributors in its launch materials. Its principal stay demonstration makes use of eight audio system, whereas its long-form recording accommodates 11 labeled contributors. The 20-plus quantity is a acknowledged mannequin functionality somewhat than the participant depend within the public demos.
At $0.18 per hour, Muse competes aggressively on worth
Meta's pricing makes the aggressive image extra attention-grabbing.
According to its Muse Voice Transcribe developer page, Muse prices $3 per 1,000 minutes, or $0.18 per hour. Streaming and non-streaming transcription price the identical, and Meta says zero-data-retention processing is priced at parity with normal processing. Billing applies to audio really processed and is rounded all the way down to complete seconds.
Standardizing publicly posted charges to 1 hour of streaming audio provides the next tough comparability:
|
Streaming speech-to-text service |
Approx. public price/hour |
Real-time diarization |
|
$0.12 |
Included; as much as 15 audio system |
|
|
$0.18 |
Included; 20+ audio system |
|
|
$0.20 |
Supported; most not acknowledged |
|
|
$0.24 |
Included; 50 default, configurable to 100 |
|
|
~$0.324 overseas |
No comparable most documented in supply reviewed |
|
|
~$0.35 base / ~$0.47 with diarization |
$0.12/hour diarization add-on |
|
|
$0.39 PAYG |
Not supported in actual time |
|
|
$0.45 base / $0.57 with diarization |
$0.12/hour add-on; as much as 10 audio system |
|
|
~$0.54 blended |
Not supported in stay mode |
|
|
~$0.60 in AWS's N. Virginia streaming instance |
Included; as much as 30 audio system |
|
|
$1.02 |
Diarization not listed as a mannequin functionality |
The comparability is essentially imperfect. Qwen's worth varies by deployment geography; its overseas real-time price of $0.00009 per second works out to about $0.324 per hour. Google's Gemini determine is an estimated blended token price somewhat than a flat hourly tariff. AWS costs differ by area and utilization tier. ElevenLabs lists $0.39 per hour on its API pricing web page however advertises $0.28 per hour or decrease on annual Business plans.
Deepgram's pricing notably illustrates why feature-level comparisons matter: its present Nova-3 Multilingual streaming price is about $0.35 per hour, however speaker diarization costs another $0.002 per minute, bringing the comparable complete to roughly $0.47 per hour. AssemblyAI equally lists $0.45 per hour for Universal-3.5 Pro Realtime and one other $0.12 per hour for streaming diarization.
Cartesia is tougher to normalize as a result of Ink-2 is packaged through monthly credit plans somewhat than a easy metered PAYG hourly price. Its $5 Pro plan consists of roughly 9 hours and 16 minutes of Ink-2 transcription, which works out to about $0.54 per transcription hour if each credit score is consumed solely on STT. That shouldn’t be handled as equal to a standalone $0.54 hourly API tariff.
Even with these caveats, Muse's positioning is obvious. It will not be absolutely the least expensive streaming transcription service — Soniox presently publishes a decrease equal price — however $0.18 per hour with diarization included places Meta towards the low finish of the market, particularly in opposition to suppliers that cost individually for speaker attribution.
At 1,000 hours of processed audio, Meta's public price implies roughly $180 in transcription fees.
Meta additionally leads its launch accuracy benchmarks
Price issues much less if it comes with a big accuracy penalty. Meta's benchmark materials argues the other.
On the Artificial Analysis AA-WER Streaming Index provided with the launch, Muse information a 3.1% final-transcription phrase error price, forward of Cartesia Ink-2 at 3.4%, ElevenLabs Scribe v2 Realtime at 3.6%, Qwen3 ASR Flash Realtime at 3.7%, GPT Live Transcribe and Grok Speech to Text Streaming at 3.9%, and Gemini 3.5 Transcribe Live and AssemblyAI U3.5 Realtime Pro at 4.0%.
Meta factors out that Muse took the primary spot on third-party impartial AI benchmarking agency Artificial Analysis' streaming speech-to-text evaluation as of September 1. Meta published the following benchmark charts in its launch post.
Its diarization end result could also be much more related to the product's positioning. Meta studies a mean 17.5% diarization error price throughout AMI-IHM, AMI-SDM and VoxConverse, decrease than the competing methods proven in its chart.
Speaker capability and diarization error price shouldn’t be conflated. A platform able to representing 100 individuals will not be robotically higher at accurately attributing speech than one supporting 20, and Meta's benchmark doesn’t take a look at each competitor working at its marketed most speaker depend.
There are deployment tradeoffs as properly. Meta's API presently offers turn-level however not word-level timestamps, and it doesn’t expose word-level confidence scores, sound-event detection or emotion detection. The documentation additionally specifies eight concurrent streams per tenant by default and real-time periods of as much as 60 minutes earlier than an software should reconnect.
Still, Muse's launch creates an unusually sharp price-performance proposition. Its 20-plus-speaker diarization doesn’t set up a planet document, however the document could be the much less essential metric. For enterprise builders, the bigger query is whether or not a service can protect speaker attribution, correct textual content and usable flip boundaries whereas an advanced real-world dialog remains to be unfolding.
At $0.18 per hour, with 20-plus-speaker diarization inside the identical real-time mannequin that presently leads Meta's provided streaming accuracy benchmarks, Muse Voice Transcribe provides enterprise groups a severe new possibility for assembly intelligence, stay transcription and voice-agent utilities — whereas placing extra stress on opponents to compete on speaker-aware accuracy and complete working price, not merely uncooked speech recognition.
