Meta Enters the Real-Time Speech-to-Text Arena with Muse Voice Transcribe, Offering Advanced Features at a Disruptive Price Point.

Meta is making a significant play in the rapidly evolving and increasingly competitive real-time speech-to-text market with the introduction of Muse Voice Transcribe. This new audio perception model, developed by Meta Superintelligence Labs, promises to deliver a comprehensive suite of capabilities, including streaming transcription, robust endpoint detection, and sophisticated speaker diarization for an impressive number of participants – over 20 – all accessible through a public API priced at an aggressive $0.18 per hour of processed audio. Unlike traditional systems that require audio to be fully recorded before processing, Muse is engineered to handle speech as it happens, offering a more immediate and dynamic transcription experience.

According to Meta’s official launch post, Muse Voice Transcribe is built to process speech in real-time, eliminating the need to wait for audio recordings to conclude. The model boasts support for lengthy audio exceeding an hour, seamless multilingual code-switching, flexible language and keyword biasing, and integrated diarization that eliminates the need for a separate post-processing pipeline. Meta has trained Muse across an extensive library of over 70 languages, with 25 of these languages having undergone thorough validation for the initial release. This broad language support positions Muse as a versatile tool for global applications.

While the capability to handle more than 20 speakers is a substantial achievement, it does not represent an absolute world record in the industry. A review of current vendor documentation reveals systems with higher published speaker ceilings. For instance, Speechmatics’ real-time transcription service claims to identify 50 speakers by default, with the potential to increase this limit to 100 speakers. 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, Muse’s strength lies in its integrated approach, combining high-capacity real-time diarization with low-latency transcription, endpointing, multilingual code-switching, and aggressive API pricing within a single, cohesive model. This unified offering is poised to be particularly impactful for enterprise developers building sophisticated applications such as meeting systems, call analytics platforms, live assistants, and ambient AI solutions, where the seamless integration of these features can be more critical than the raw maximum speaker count.

The increasing importance of speaker diarization is fundamentally reshaping the core of voice technology stacks. Historically, speech recognition systems primarily focused on answering the question, "What was said?" Diarization adds a crucial layer of intelligence by addressing the question, "Who said it?" This distinction is paramount as transcribed audio feeds into downstream AI systems. Without accurate speaker attribution, even a perfectly transcribed meeting can lead to an unreliable corporate record, potentially misattributing approvals, commitments, or objections to the wrong individuals. This challenge extends to customer service analytics, compliance workflows, and AI agents operating in environments where multiple individuals participate in conversations.

Muse addresses this by incorporating speaker attribution directly into its autoregressive multimodal architecture. Meta explains that audio is processed in small 80-millisecond chunks, translating to 12.5 chunks per second, with each chunk converted into a soft token. At each step, the model dynamically decides whether to consume more audio or emit text. This mechanism, termed "adaptive delay" by Meta, allows Muse to deviate from a fixed latency budget for every word. Instead, it can afford to wait longer when speech is ambiguous, gathering more context, and commit to emitting text earlier when sufficient information is available. Meta leverages reinforcement learning to train this adaptive behavior, rewarding both word-error-rate reduction and timely delivery. The technical intricacies of Muse’s architecture are further detailed in Meta’s research publication on the subject.

Within this architecture, speaker attribution and endpointing are treated as integral components of the same token sequence. Special tokens, such as <|start_of_turn|>, are used to mark potential new speaker turns, while tokens like <|speaker_A|> identify the specific speaker. Separate onset and endpoint tokens delineate the boundaries of speech segments. Meta’s approach of training Automatic Speech Recognition (ASR), diarization, and endpointing collaboratively, rather than treating speaker clustering as a separate downstream process, is a key differentiator. Meta’s Model API documentation for speech-to-text further emphasizes this integrated approach, exposing diarization as a first-class operating mode alongside push-to-talk and endpointing functionalities. Speaker labels, such as "A" and "B," are session-scoped rather than tied to verified identities, and the API provides turn-level timestamps, offering a more consolidated view of conversational flow.

While Muse’s capacity for handling over 20 speakers is notable, a detailed comparison with competitors reveals varying approaches and published limits in speaker diarization. Speechmatics stands out with its explicit real-time capacity claim, stating that its system supports 50 speakers by default and can be configured to handle up to 100. Amazon Transcribe also surpasses Meta’s stated figure, with the capability to differentiate up to 30 unique speakers and providing specific guidance for speaker partitioning in streaming transcription. Soniox supports diarization in both real-time and asynchronous modes but documents a maximum of 15 speakers per session. AssemblyAI’s streaming diarization system allows developers to set a max_speakers parameter between one and ten. Both Soniox and AssemblyAI acknowledge the inherent challenges of real-time speaker attribution, noting that streaming systems must make decisions with less future audio context compared to offline models. xAI’s current Speech-to-Text API also includes speaker diarization in its streaming mode, but its documentation, as reviewed for this analysis, does not publish a maximum speaker count, making a direct ceiling comparison with Muse impossible. Consequently, it would be inaccurate to present Muse’s 20-plus speaker capability as a new global record; that distinction, based on explicitly documented real-time capacity, appears to belong to Speechmatics with its configurable 100-speaker ceiling.

Furthermore, Meta’s launch materials do not showcase demonstrations with over 20 simultaneous participants. Their primary 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 observed in their public demonstrations.

The competitive landscape for real-time speech-to-text services becomes even more intriguing when considering Meta’s pricing strategy. Muse Voice Transcribe is priced at $3 per 1,000 minutes, which translates to $0.18 per hour. This pricing applies equally to both streaming and non-streaming transcription, and Meta further states that zero-data-retention processing is offered at the same rate as standard processing. Billing is based on the actual audio processed and is rounded down to whole seconds.

To provide a clearer picture of Muse’s competitive positioning on price, a standardization of publicly posted rates to one hour of streaming audio reveals the following approximate costs:

It is important to note that this comparison is inherently imperfect due to variations in vendor pricing models and feature sets. For example, Qwen’s pricing varies geographically, with its international real-time rate translating to approximately $0.324 per hour. Google’s Gemini figure is an estimated blended token cost, not a flat hourly tariff. AWS pricing is subject to regional and usage-tier variations. Deepgram’s pricing illustrates the impact of feature-level charges, where its Nova-3 Multilingual streaming rate of approximately $0.35 per hour, when combined with the $0.002 per minute diarization add-on, reaches roughly $0.47 per hour. AssemblyAI follows a similar pattern, with its Universal-3.5 Pro Realtime priced at $0.45 per hour and an additional $0.12 per hour for streaming diarization. Cartesia’s Ink-2 is packaged in monthly credit plans, making direct hourly comparisons challenging.

Despite these caveats, Meta’s pricing positions Muse Voice Transcribe very favorably. While not the absolute cheapest streaming transcription service – Soniox currently offers a lower equivalent rate – the $0.18 per hour price point, which includes comprehensive diarization, places Meta at the lower end of the market, particularly when compared to providers that charge extra for speaker attribution. For an enterprise processing 1,000 hours of audio, Meta’s public rate suggests transcription costs of approximately $180.

Beyond its aggressive pricing, Meta also claims superior performance in its launch accuracy benchmarks. On the Artificial Analysis AA-WER Streaming Index, Muse reportedly achieves a final-transcription word error rate of 3.1%, outperforming 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 (3.9%), and Gemini 3.5 Transcribe Live and AssemblyAI U3.5 Realtime Pro (4.0%). Meta highlights that Muse secured the top position on Artificial Analysis’ independent streaming speech-to-text evaluation as of September 1st.

The diarization performance presented by Meta may be 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 that of competing systems depicted in their benchmark charts.

It is crucial to understand that speaker capacity and diarization error rate are not directly interchangeable metrics. A platform capable of identifying 100 speakers is not automatically superior in accurately attributing speech to those speakers compared to one supporting 20, and Meta’s benchmark does not encompass testing against every competitor at their maximum advertised speaker counts.

Furthermore, there are practical deployment considerations to factor in. 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. The documentation also specifies eight concurrent streams per tenant by default and real-time sessions lasting up to 60 minutes before an application needs to re-establish a connection.

Nevertheless, Muse’s launch introduces a compelling value proposition, characterized by a sharp balance of price and performance. While its 20-plus speaker diarization capability may not set a new world record, the record itself might be a less significant factor for many enterprise users. The more pertinent question for businesses is whether a service can consistently maintain accurate speaker attribution, precise transcription, and usable conversational turn boundaries in dynamic, real-world discussions. At $0.18 per hour, with integrated diarization for over 20 speakers, and leading Meta’s own streaming accuracy benchmarks, Muse Voice Transcribe presents a formidable new option for enterprise teams seeking advanced capabilities in meeting intelligence, live transcription, and voice-agent infrastructure. This offering is expected to exert significant pressure on competitors to enhance their speaker-aware accuracy and overall cost-effectiveness, moving beyond solely focusing on raw speech recognition performance.