Meta Enters Real-Time Speech-to-Text Arena with Muse Voice Transcribe, Offering Integrated Diarization and Aggressive Pricing

Meta is making a significant move into the increasingly crowded real-time speech-to-text market with the introduction of Muse Voice Transcribe. This novel audio perception model, developed by Meta Superintelligence Labs, distinguishes itself by seamlessly integrating streaming transcription, precise endpoint detection, and robust speaker diarization for conversations involving over 20 participants. The company has set a compelling public API price of just $0.18 per hour of processed audio, positioning Muse as a formidable contender in a landscape where nuanced understanding of spoken interactions is paramount.

Muse is engineered for true real-time processing, a departure from traditional models that often require audio to be fully recorded before transcription can commence. Meta’s official launch announcement highlights Muse Voice Transcribe’s capabilities, emphasizing its support for extended audio sessions exceeding an hour, fluid multilingual code-switching, granular language and keyword biasing, and crucially, integrated diarization that eliminates the need for a separate post-processing pipeline. The model’s extensive training across more than 70 languages, with 25 rigorously validated for its initial release, underscores Meta’s commitment to global applicability and nuanced linguistic understanding.

While the advertised capability of handling over 20 speakers is substantial, it is important to note that this figure does not represent a new industry record. A comparative review of existing vendor documentation reveals systems with higher published speaker ceilings. For instance, Speechmatics’ real-time transcription service boasts the ability to identify up to 50 speakers by default, with an option to increase this to 100. Similarly, Amazon Transcribe’s diarization documentation specifies a maximum of 30 unique speakers, even for streaming transcription scenarios. Despite these higher individual speaker counts from competitors, Meta’s Muse Voice Transcribe positions itself towards the higher end of the market not just in terms of speaker capacity, but through its holistic approach. The broader proposition of Muse lies in its ability to combine high-capacity real-time diarization with low-latency transcription, sophisticated endpointing, seamless multilingual code-switching, and aggressive API pricing, all within a single, unified model. This integrated approach is particularly relevant for enterprise developers building complex applications such as advanced meeting systems, in-depth call analytics platforms, intelligent live assistants, and ambient AI solutions, where the quality of speaker attribution and real-time understanding can be more critical than the absolute maximum speaker count.

The integration of speaker diarization into the core voice processing stack represents a significant evolution in how speech technologies are being developed and utilized. Traditionally, speech recognition systems primarily answered the question: "What was said?" Diarization fundamentally adds another critical dimension: "Who said it?" This distinction is not merely an academic one; it becomes critically important as transcribed audio feeds into downstream AI systems. For example, a meeting assistant that can accurately transcribe every sentence might still produce an unreliable corporate record if it incorrectly attributes an approval, a commitment, or an objection to the wrong participant. The implications are far-reaching, impacting customer-service analytics, compliance workflows, and AI agents operating in environments where multiple individuals are actively participating in conversations.

Muse addresses this challenge by incorporating speaker attribution directly into its autoregressive multimodal architecture. Meta explains that audio data is processed in small 80-millisecond chunks, translating to 12.5 chunks per second. Each chunk is transformed into a "soft token," and at each step, the model dynamically decides whether to ingest more audio or emit transcribed text. Meta refers to this innovative mechanism as "adaptive delay." Unlike systems that impose a uniform latency budget for every word, Muse can afford to wait longer for more context when speech is ambiguous, and conversely, commit to emitting text earlier when sufficient context has been acquired. Meta further elaborates that reinforcement learning plays a key role in training this adaptive behavior, balancing rewards for both word-error rate and acceptable delays. A deeper technical dive into Muse’s architecture is available in Meta’s research publications.

Crucially, speaker attribution and endpointing are treated as integral components of the same token sequence within Muse’s architecture. Specific tokens, such as <|start_of_turn|>, signal a potential new speaker turn, while tokens like <|speaker_A|> are used to identify the speaker. Separate onset and endpoint tokens are employed to delineate the precise boundaries of spoken segments. Meta emphasizes that its training methodology involves co-training Automatic Speech Recognition (ASR), diarization, and endpointing together, rather than treating speaker clustering as an independent, downstream process. This unified approach is expected to yield more accurate and coherent results. Meta’s Model API speech-to-text documentation further reinforces this integrated design by exposing diarization as a first-class operating mode, alongside traditional push-to-talk and endpointing functionalities. The API provides speaker labels (e.g., A and B) that are scoped to a particular session, rather than attempting to infer verified identities, and offers turn-level timestamps, which are generally more practical for dialogue analysis than word-level timestamps in real-time scenarios.

While Meta’s claim of supporting over 20 speakers is a significant advancement, a closer look at the competitive landscape reveals that this figure is not an absolute industry peak. Careful comparison of speaker-count capabilities requires attention to the differing implementation strategies and varying levels of public disclosure among vendors. Speechmatics currently holds the strongest explicit claim for real-time capacity identified in this review. Their real-time STT documentation indicates that speaker diarization is available live, and their real-time FAQ states that the system supports 50 speakers by default, with the possibility of increasing this to 100. Amazon Web Services (AWS) also surpasses Meta’s stated figure, with Amazon Transcribe capable of differentiating a maximum of 30 unique speakers, and AWS provides explicit guidance for speaker partitioning within streaming transcription.

Other notable players in the market also offer speaker diarization, albeit with different limitations. Soniox supports diarization in both real-time and asynchronous processing but documents a maximum of 15 speakers per session. AssemblyAI’s streaming diarization system allows developers to set the max_speakers parameter between one and 10. Both Soniox and AssemblyAI acknowledge the inherent difficulty of live speaker attribution in streaming systems, which must make decisions with less future audio context compared to offline models. xAI’s current Speech-to-Text API also supports speaker diarization in streaming mode, but its documentation, as reviewed for this report, does not publicly state a maximum diarized-speaker count, making a direct ceiling comparison with Muse impossible. Therefore, it would be inaccurate to position Muse’s "20-plus" speaker capability as a new global record. The highest explicitly documented real-time number identified in this survey remains Speechmatics’ configurable 100-speaker ceiling. Furthermore, Meta’s own launch materials do not showcase demonstrations with more than 20 simultaneous participants; their principal live demonstration features eight speakers, and a long-form recording includes 11 labeled participants. The "20-plus" figure represents a stated model capability rather than the participant count in their public demonstrations.

The pricing strategy employed by Meta significantly alters the competitive dynamics of the speech-to-text market. Muse Voice Transcribe is priced at $3 per 1,000 minutes, which translates to an attractive $0.18 per hour. This pricing is consistent for both streaming and non-streaming transcription, and Meta clarifies that zero-data-retention processing is offered at the same parity price. Billing is applied to audio that is actually processed and is rounded down to whole seconds, offering a predictable cost structure.

To provide a clearer comparison of real-time streaming speech-to-text services, a standardization of publicly posted rates to one hour of streaming audio can be performed, yielding the following approximate cost comparison:

It is important to acknowledge the inherent imperfections in such comparisons. Qwen’s pricing, for instance, varies by geographical deployment, with its international real-time rate of $0.00009 per second translating to approximately $0.324 per hour. Google’s Gemini figure represents an estimated blended token cost rather than a fixed hourly tariff. AWS pricing also exhibits variability based on region and usage tier. Deepgram’s pricing model is particularly illustrative of the importance of feature-level comparisons: its Nova-3 Multilingual streaming rate is around $0.35 per hour, but speaker diarization incurs an additional cost of $0.002 per minute, bringing the total comparable cost to approximately $0.47 per hour. AssemblyAI follows a similar pattern, listing $0.45 per hour for Universal-3.5 Pro Realtime with an additional $0.12 per hour for streaming diarization. Cartesia’s Ink-2 pricing is more complex to normalize, as it is packaged through monthly credit plans rather than a straightforward metered PAYG hourly rate. Their $5 Pro plan includes approximately nine hours and 16 minutes of Ink-2 transcription, equating to about $0.54 per transcription hour if all credits are used solely for STT, which should not be directly equated to a standalone $0.54 hourly API tariff.

Despite these caveats, Muse’s pricing positioning is unequivocally clear. While it may not be the absolute cheapest streaming transcription service – Soniox currently publishes a lower equivalent rate – the $0.18 per hour price point, which includes sophisticated diarization, firmly places Meta towards the lower end of the market, especially when contrasted with providers that levy separate charges for speaker attribution. At a scale of 1,000 hours of processed audio, Meta’s public rate implies transcription charges of approximately $180, a highly competitive figure for enterprise-level applications.

Beyond its aggressive pricing, Meta also claims leadership in accuracy benchmarks. On the Artificial Analysis AA-WER Streaming Index, which was supplied alongside the product launch, Muse achieved a final-transcription word error rate of 3.1%. This performance places it ahead of competitors such as Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Realtime (3.6%), Qwen3 ASR Flash Realtime (3.7%), GPT Live Transcribe and Grok Speech to Text Streaming (both at 3.9%), and Gemini 3.5 Transcribe Live and AssemblyAI U3.5 Realtime Pro (both at 4.0%). Meta highlights that Muse secured the top position on the independent AI benchmarking firm Artificial Analysis’ streaming speech-to-text evaluation as of September 1.

The diarization performance reported by Meta is perhaps even more critical to the product’s market positioning. The company reports an average diarization error rate of 17.5% across the AMI-IHM, AMI-SDM, and VoxConverse datasets, which is lower than the competing systems presented in their comparative chart. It is important to emphasize that speaker capacity and diarization error rate are not directly interchangeable metrics. A platform capable of representing 100 individuals does not automatically imply superior accuracy in attributing speech compared to one supporting 20. Furthermore, Meta’s benchmark analysis does not encompass every competitor operating at their respective maximum speaker counts.

Practical deployment considerations also come into play. Meta’s API currently provides turn-level timestamps but not word-level timestamps, and it does not expose word-level confidence scores, sound-event detection, or emotion detection capabilities. The documentation also specifies a default of eight concurrent streams per tenant and real-time sessions limited to 60 minutes before an application must re-establish a connection. Nevertheless, the launch of Muse Voice Transcribe presents an exceptionally sharp price-performance proposition. While its 20-plus-speaker diarization may not set a new world record, the record itself might be a less significant metric than the overall value delivered. For enterprise developers grappling with the complexities of real-world conversations, the more pertinent question is whether a service can consistently preserve speaker attribution, maintain transcription accuracy, and provide usable turn boundaries while a dynamic dialogue is actively unfolding. At $0.18 per hour, with 20-plus-speaker diarization integrated into a real-time model that Meta claims leads its own streaming accuracy benchmarks, Muse Voice Transcribe offers enterprise teams a compelling new option for enhancing meeting intelligence, facilitating live transcription, and building robust voice-agent infrastructure. This launch also intensifies pressure on competitors to focus not only on raw speech recognition capabilities but also on speaker-aware accuracy and overall operational cost.