Meta has entered the competitive real-time speech-to-text market with Muse Voice Transcribe, a speech recognition model that combines live transcription, endpoint detection, and speaker diarization for more than 20 speakers. The public API is priced at just $0.18 per hour of processed audio, positioning Muse as a low-cost option for enterprise voice applications.
Developed by Meta Superintelligence Labs, Muse Voice Transcribe processes audio as it is spoken instead of waiting for a recording to finish. According to Meta’s Muse Voice Transcribe announcement, the model supports audio longer than one hour, multilingual code-switching, language and keyword biasing, and integrated speaker diarization without requiring a separate post-processing pipeline. Muse was trained on more than 70 languages, with 25 languages extensively validated for the initial release.
Support for more than 20 speakers is substantial, although it is not a record for real-time speech recognition. Speechmatics’ real-time transcription service states that it can identify 50 speakers by default and up to 100 speakers with an increased limit. Amazon Transcribe’s diarization documentation specifies support for up to 30 unique speakers, including in streaming transcription.
Even so, Muse sits toward the higher end of the market when its full feature set is considered. Meta is combining high-capacity real-time diarization with low-latency transcription, endpoint detection, multilingual code-switching, and aggressive API pricing in a single model.
For enterprise developers building video conferencing platforms, call center analytics tools, live AI assistants, meeting applications, and ambient AI systems, that combination may matter more than the maximum number of supported speakers alone.
Why real-time speaker diarization matters
Traditional automatic speech recognition answers a relatively straightforward question: What was said? Speaker diarization adds a second question: Who said it?
This distinction becomes critical when transcripts are used by downstream AI systems. A meeting assistant may transcribe every sentence accurately, but it can still produce an unreliable business record if an approval, commitment, or objection is attributed to the wrong participant. The same risk applies to customer service analytics, compliance workflows, sales intelligence, and AI agents operating in rooms where multiple people can speak.
Muse incorporates speaker information directly into an autoregressive multimodal architecture. Meta says audio is processed in 80-millisecond chunks, equivalent to 12.5 chunks per second, with each chunk converted into a soft token. At each step, the model decides whether to process more audio or generate text.
Meta refers to this approach as adaptive delay. Instead of applying the same latency target to every word, Muse can wait longer when speech is ambiguous and commit more quickly when enough context is available. According to Meta, reinforcement learning combines word error rate with delayed rewards to train this behavior. The company provides a detailed overview in its technical explanation of Muse Voice Transcribe.
Speaker attributes and endpoint information are included in the same token sequence. A <|start_of_turn|> token marks the beginning of a potential new speaker turn, while tokens such as <|speaker_A|> identify participants. Separate start and end tokens help the model identify speech boundaries.
Meta says it trains automatic speech recognition, diarization, and endpoint detection together rather than treating speaker clustering as an unrelated downstream process. The company’s Muse Voice Transcribe API documentation exposes diarization as a core operating mode alongside push-to-talk and endpointing.
Speaker labels such as A and B are scoped to the active session and do not represent verified identities. The API currently provides timestamps at the turn level rather than for every individual word.
How Muse compares with other real-time diarization APIs
Speaker-count comparisons require caution because vendors implement diarization differently, and many providers do not publish a maximum number of speakers. A higher limit also does not necessarily mean better speaker attribution or transcription accuracy.
Speechmatics’ real-time speech-to-text documentation makes some of the strongest explicit capacity claims identified in this comparison. Its real-time FAQ states that the platform supports 50 speakers by default and can be configured for up to 100 speakers.
AWS also lists a higher maximum than Meta. Amazon Transcribe can distinguish up to 30 unique speakers, and AWS provides instructions for enabling speaker identification during streaming transcription.
Soniox supports diarization for both real-time and asynchronous processing, but its documentation lists support for up to 15 speakers per session. AssemblyAI’s streaming diarization system lets developers configure max_speakers. The company notes that live speaker identification is more challenging than offline diarization because streaming systems must make decisions without access to future audio context.
xAI’s Speech-to-Text API also supports speaker diarization in streaming mode. However, the documentation reviewed for this article does not publish a maximum speaker count, so its limit cannot be directly compared with Muse.
As a result, it would be inaccurate to describe Muse’s 20-plus speaker capability as a new world record. The highest explicitly documented configurable real-time limit identified here is Speechmatics’ support for up to 100 speakers.
Meta also does not demonstrate more than 20 simultaneous participants in its public presentation materials. The main live demonstration includes eight speakers, while an extended recording contains 11 labeled participants. The 20-plus figure represents the model’s stated capability, not the number of people shown in the public demo.
Muse Voice Transcribe pricing starts at $0.18 per hour
Meta’s pricing makes Muse Voice Transcribe particularly competitive for developers processing large volumes of audio.
According to the Muse Voice Transcribe developer page, the API costs $3 per 1,000 minutes, equivalent to $0.18 per hour. Meta says streaming and non-streaming transcription use the same pricing, while zero-data-retention processing is priced at the same rate as standard processing. Billing is based on the actual amount of audio processed and is rounded down to the second.
Published prices are difficult to compare directly because providers use different billing models, regions, features, and usage tiers. Qwen pricing varies by deployment region. Its international real-time rate is listed at $0.00009 per second, or approximately $0.324 per hour. Google Gemini pricing is based on estimated blended token usage rather than a flat hourly rate. AWS pricing varies by region and usage volume. ElevenLabs lists an API price of approximately $0.39 per hour, while annual business plans advertise a lower effective rate.
Deepgram’s pricing demonstrates why feature-level comparisons are important. The company’s Nova-3 multilingual streaming transcription costs approximately $0.35 per hour, while speaker diarization adds $0.002 per minute. With diarization included, the estimated total is approximately $0.47 per hour. AssemblyAI charges approximately $0.45 per hour for Universal-3.5 Pro Realtime and an additional $0.12 per hour for streaming diarization.
Cartesia packages Ink-2 through monthly credit plans rather than offering a directly comparable pay-as-you-go hourly rate. Its $5 Pro plan includes approximately nine hours and 16 minutes of Ink-2 transcription, which equals roughly $0.54 per hour if all credits are used exclusively for speech-to-text. That figure should not be treated as equivalent to a standalone $0.54-per-hour API price.
Despite these differences, Muse’s market position is clear. It is not the cheapest streaming transcription service in every comparison, and Soniox publishes a lower comparable rate in some cases. However, at $0.18 per hour with speaker diarization included, Muse is positioned toward the lower end of the market, particularly against providers that charge separately for speaker attribution.
At Meta’s published rate, processing 1,000 hours of audio would cost approximately $180 before any applicable usage or account-specific considerations.
Meta reports strong real-time transcription accuracy
Low pricing matters less if it comes with a significant accuracy penalty. Meta’s launch benchmarks suggest that Muse is competitive on both cost and transcription quality.
In the Artificial Analysis streaming speech-to-text evaluation, Muse recorded a final word error rate of 3.1%, according to Meta’s launch materials. The comparison results listed Cartesia Ink-2 at 3.4%, Celeb-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 Universal-3.5 Realtime Pro at 4.0%.
Meta says Muse ranked first among the systems evaluated by Artificial Analysis’ streaming speech-to-text benchmark as of September 1. The company published the benchmark charts in its Muse Voice Transcribe launch post.
Muse’s diarization results may further strengthen its product positioning. Meta reports an average diarization error rate of 17.5% across the AMI-IHM, AMI-SDM, and VoxConverse datasets, which the company says is lower than the competing systems shown in its comparison chart.
Speaker capacity and diarization error rate measure different things. A service that can represent 100 speakers will not automatically attribute speech more accurately than a service supporting 20 speakers. Meta’s benchmark results also do not test every competing system at its maximum advertised speaker count.
Muse has other implementation trade-offs. The API currently provides turn-level timestamps but not word-level timestamps. Meta also does not list word-level confidence scores, sound-event detection, or emotion detection among the documented features. The documentation specifies eight concurrent streams per tenant by default and real-time sessions of up to 60 minutes before an application must reconnect.
What Meta’s Muse launch means for enterprise developers
Muse Voice Transcribe does not set a new record for the largest number of speakers supported in real-time transcription. However, raw speaker capacity may be less important than the overall balance of accuracy, latency, diarization, multilingual support, and operating cost.
At $0.18 per hour, Meta is offering integrated diarization for more than 20 speakers within a real-time speech-to-text model that the company says leads current streaming accuracy benchmarks. That makes Muse a serious option for meeting intelligence, live transcription, customer service analytics, and voice-agent infrastructure.
The launch also puts additional pressure on competitors to improve speaker-aware accuracy and reduce total operating costs rather than competing only on basic speech recognition or maximum speaker counts.
Source: venturebeat.com