Meta Releases Muse Real-time Voice Transcription Model at $0.18 per Hour, Training Covers Over 70 Languages
Meta Superintelligence Labs has released Muse Voice Transcribe, integrating real-time speech recognition, speaker separation, and utterance end detection into a single model, natively supporting over 1 hour of audio and conversations with more than 20 speakers without additional post-processing.
In the English streaming transcription benchmark by Artificial Analysis, Muse Voice Transcribe achieved a final text word error rate of 3.1%, ranking first; GPT Live Transcribe had 3.9%, and Gemini 3.5 Transcribe Live had 4.0%. A lower word error rate indicates that the model's output has a smaller word-level difference from the human reference text.
Muse outputs the final text about 0.16 seconds after the user stops speaking. Traditional real-time transcription often requires a preset fixed waiting window; a shorter window is faster but may truncate sentences, while a longer window is accurate but increases delay. Muse uses 80-millisecond audio chunks as input, allowing the model to autonomously decide whether to continue listening or end the current utterance and submit the transcription result.
Meta claims it optimizes transcription accuracy and endpoint detection delay simultaneously through reinforcement learning. Simple, contextually clear words can be confirmed earlier; ambiguous words that require subsequent audio to distinguish will have more listening time retained. This changes endpoint detection from fixed rules to dynamically allocating computation and waiting budgets based on speech content.
The model training covers over 70 languages, with 25 of them validated at launch, including Mandarin, English, Cantonese, Japanese, Korean, Spanish, Arabic, Hindi, Thai, Vietnamese, and Portuguese. Muse also natively supports code-switching, recognizing mixed languages such as Chinese and English within or between sentences.
Muse is integrated with Meta Model API, Meta AI for Mac, and Muse Code. The API model is muse-voice-transcribe-1.0, priced at $3 per 1,000 audio minutes, or $0.18 per hour; it is currently a managed API service, and Meta has not opened the model weights.
In the market mechanism, buyers include enterprises and developers building meeting notes, real-time customer service, voice agents, interview transcriptions, and multilingual content tools; sellers include voice infrastructure providers such as Meta, OpenAI, Google, ElevenLabs, Deepgram, AssemblyAI, and Cartesia. Funding will flow from manual transcription, post-processing labeling, and multi-model stitching pipelines to low-latency voice models, API calls, meeting intelligence, customer service automation, and voice agents; platforms that can deliver accurate transcription, speaker identity, and real-time endpoint detection will benefit, while providers relying on manual shorthand or high-latency batch processing will face pressure.
Source: Public Information
ABAB AI Insight
The key aspect of Meta's release of Muse Voice Transcribe is not just the addition of voice transcription capabilities, but the first time it is referred to as a real-time audio perception model from Meta Superintelligence Labs. In the past, voice applications often combined ASR, VAD, speaker separation, punctuation recovery, and post-processing into a pipeline: one model converts sound to text, another module determines pauses, and a third module identifies different speakers. Muse processes these three core tasks within a single autoregressive multimodal model, reducing delays, error accumulation, and engineering integration complexity caused by cross-model transfers.
From a capital perspective, the API price of $0.18 per hour shifts voice transcription from manual services to cloud infrastructure billed by the minute. Enterprises send audio streams to the Meta API, and Meta generates revenue from model inference, GPU clusters, real-time streaming scheduling, and language optimization; customers shift budgets originally allocated for manual transcription, quality control in customer service, meeting minutes, and recording post-processing to API, product integration, and subsequent workflow automation. Muse's integration with Meta AI for Mac and Muse Code indicates that Meta is not only selling voice capabilities but also establishing distribution channels and internal usage through its own desktop AI and development tools.
Historically, Twilio encapsulated phone and SMS infrastructure as APIs, Stripe encapsulated payment capabilities as APIs, and OpenAI Whisper made high-quality voice transcription a foundational model capability for developers. Muse's differentiation lies in shifting the competitive focus to "real-time, multilingual, multi-speaker, and low-latency" scenarios: meeting assistants need to identify who is speaking, voice agents need to quickly determine if the user has finished speaking, and bilingual families or global customer service need to understand intra-sentence language switching. ElevenLabs, Deepgram, and AssemblyAI have already laid out in the real-time voice market, and Meta enters with a benchmark WER of 3.1% and a final output delay of 0.16 seconds, positioned at the stage of upgrading from general transcription to real-time interactive infrastructure.
This represents a technological substitution. What is being replaced is not human language itself, but the waiting, manual recording, and multi-module processing required before voice enters software systems. In the past, recording a multi-speaker meeting required completing transcription, identifying speakers, and manually organizing action items; when the model can stream recognize, label speakers, and confirm completion within 0.16 seconds, voice can directly serve as real-time input for customer service systems, coding tools, meeting software, and AI agents. The real new barrier is not just low WER, but seamlessly integrating audio perception into the next steps: generating minutes, updating CRM, triggering workflows, writing code, or continuing conversations with voice agents.
ABAB News · Cognitive Laws
Turning speech into text, delay determines interaction value
Accurately identifying content and distinguishing who is speaking is more valuable
Unified perception by the model allows workflows to be executed automatically