{"slug": "metas-muse-voice-transcribe-handles-real-time-speech", "title": "Meta’s Muse Voice Transcribe handles real-time speech", "summary": "Meta is releasing Muse Voice Transcribe, a real-time speech model that combines streaming transcription, speaker diarization, and endpointing, capable of handling 20 or more speakers and code-switching, with API access priced at $3 per 1,000 audio minutes. The model, developed by Meta Superintelligence Labs, processes audio in 80-millisecond chunks and uses adaptive delay to balance speed and accuracy, and is already powering dictation in Meta's desktop Mac app and Muse Code.", "body_md": "[AI](/ai/)\n\n• 5 min read\n\n# Meta’s Muse Voice Transcribe handles real-time speech\n\nMeta’s Muse Voice Transcribe handles streaming speech, 20+ speakers and code-switching, with API access priced at $3 per 1,000 audio minutes.\n\nImage: [Engadget](https://www.engadget.com/2249112/meta-new-ai-transcription-model-can-distinguist-between-multiple-speakers-and-languages-in-real-time/)\n\nMeta is releasing **Muse Voice Transcribe**, a real-time speech model that combines streaming transcription, speaker diarization and endpointing in one system. The model can process conversations with **20 or more speakers**, switch between languages mid-sentence and transcribe audio sessions lasting more than an hour.\n\nThe release is the first real-time audio perception model from Meta Superintelligence Labs. Meta says it ranks first on Artificial Analysis for streaming speech-to-text and on public diarization benchmarks, although those rankings are stated by Meta and the supplied material does not include independent benchmark results or test conditions.\n\nMuse Voice Transcribe is already powering dictation in Meta’s desktop Mac app and in Muse Code. Developers can also access it through Meta’s Model API, priced at **$3 for 1,000 audio minutes**. Meta has also published a demo through its research blog.\n\n## How the streaming model decides when to transcribe\n\nThe implementation differs from a conventional speech-to-text pipeline that records audio, segments it and processes the segments afterward. As described in [Meta’s technical release for Muse Voice Transcribe](https://research.meta.ai/blog/introducing-muse-voice-transcribe), the model is an autoregressive multimodal system from the Muse Spark family.\n\nAudio arrives in **80-millisecond chunks**, or **12.5 chunks per second**. Each chunk is converted into one soft token. At every step, the model chooses between continuing to listen and emitting a text token. To keep listening, it predicts a special `<|next_audio|>`\n\ntoken, which is then replaced by the next audio chunk. When the stream ends, an `<|empty_audio|>`\n\ntoken tells the model that no more audio is available, after which it emits the remaining text without requesting another audio chunk.\n\nThat design gives the model direct control over transcription delay. Waiting for more audio can improve recognition of difficult words, but it also increases latency. Muse Voice Transcribe uses **adaptive delay**, changing how long it waits on a word-by-word basis. Meta says it trained that behavior with reinforcement learning, combining word-error-rate and delay rewards multiplicatively, and that the resulting system reaches the Pareto front for the tradeoff between speed and final-transcription accuracy.\n\nThe release does not provide the actual latency values, word-error rates or hardware requirements behind that claim. “Real time” therefore describes the operating mode, not a quantified response-time guarantee.\n\n## Diarization and endpointing are built into the same model\n\nDiarization is implemented with additional special tokens rather than a separate post-processing stage. A `<|start_of_turn|>`\n\ntoken marks a possible speaker change, while a `<|speaker_A-Z|>`\n\ntag identifies the speaker. The model can predict the turn boundary as soon as a switch occurs and delay the speaker label until the end of the current audio chunk. If one person’s speech is split into multiple segments, those segments can retain the same speaker tag.\n\nEndpointing uses `<|speech_onset|>`\n\nto mark the beginning of speech and `<|speech_endpoint|>`\n\nto mark when a person has finished. Meta trains diarization and endpointing alongside streaming ASR, adding task-specific rewards to the transcription objective. For voice interfaces, the application needs to identify what was said, who said it and when a turn has ended before it can respond.\n\nMeta’s demonstrations include an eight-speaker read-through and a conversation lasting **1:00:52**, with speakers labeled A through K in the long-form example. The company says the system handles overlapping, interrupted and informal speech without required post-processing, but the supplied release does not quantify accuracy under those conditions.\n\n## Language coverage includes code-switching\n\nMuse Voice Transcribe was trained on **70-plus languages**, with **25 extensively verified** for the initial release. Meta recommends those 25 validated languages, while saying additional languages are supported. The model is designed to recognize arbitrary code-switching both within a sentence and between sentences, rather than forcing a conversation into one selected language.\n\nMeta’s examples include Mandarin-English speech that mixes technical terms and product names in the same sentence. The system also supports language, keyword and context biasing, which can improve recognition of names, places and domain-specific vocabulary. That feature could be useful for dictation, but it also creates a dependency on the context supplied by the application; Meta has not detailed the API format or privacy controls for contacts, keywords or other contextual data in the supplied material.\n\nThe first validated languages are not listed in the release text provided here. For US users, the release does not say whether a particular bilingual workflow is covered by the validated set or falls under Meta’s broader “additional languages” support.\n\n## Availability and competitive pressure\n\nFor US users, the model is available through Meta’s Mac app, where it powers dictation features across other applications. Developers can use Muse Code or the Model API, but Meta has not announced a broader integration into Facebook, Instagram, WhatsApp or other flagship services.\n\nGoogle introduced Gemini 3.5 Transcribe less than a week earlier with similar real-time, multilingual and multi-speaker capabilities. Google is expected to integrate its model into Android and eventually Chrome; Meta’s current announcement establishes availability in its Mac software and developer tooling but does not describe a comparable mobile or browser rollout.\n\nThe API price is straightforward—**$3 per 1,000 audio minutes**—but the release does not specify whether billing differs by input language, diarization, context biasing or long-form sessions. It also does not state whether Mac-app usage is capped, metered separately or included in an existing subscription.\n\nMeta’s recent product sequence has included its first dedicated coding agent, the open-weight Muse Glimmer model and the Meta AI Mac app. Muse Voice Transcribe adds audio to that sequence, with an architecture that combines speech recognition, turn detection and speaker identity in one streaming interface. Meta has not yet shown whether its benchmark lead translates into lower end-to-end latency and fewer errors than Google’s competing system in real, mixed-language conversations.\n\n## Frequently asked questions\n\n## How much does Muse Voice Transcribe cost?+\n\nMeta prices access through its Model API at $3 for 1,000 audio minutes. The company has not specified whether features such as diarization or context biasing change the price.\n\n## Where can I use Muse Voice Transcribe?+\n\nIt powers dictation in Meta’s Mac app and is available to developers through Muse Code and Meta’s Model API. Meta has not announced integrations into its flagship social apps.\n\n## How many languages does Muse Voice Transcribe support?+\n\nThe model was trained on more than 70 languages, with 25 extensively verified for the initial release. Meta recommends trying those 25 validated languages.\n\n## Can Muse Voice Transcribe identify different speakers?+\n\nYes. Meta says it supports diarization for conversations with more than 20 speakers and performs that task natively with streaming transcription.\n\n[Ava Chen](/authors/ava-chen/)\n\nAI Editor\n\nAva covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.", "url": "https://wpnews.pro/news/metas-muse-voice-transcribe-handles-real-time-speech", "canonical_source": "https://forgeeks.net/meta-muse-voice-transcribe-release/", "published_at": "2026-09-02 00:39:57+00:00", "updated_at": "2026-09-02 00:52:08.200207+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "natural-language-processing", "ai-products", "ai-research"], "entities": ["Meta", "Muse Voice Transcribe", "Meta Superintelligence Labs", "Muse Code", "Artificial Analysis"], "alternates": {"html": "https://wpnews.pro/news/metas-muse-voice-transcribe-handles-real-time-speech", "markdown": "https://wpnews.pro/news/metas-muse-voice-transcribe-handles-real-time-speech.md", "text": "https://wpnews.pro/news/metas-muse-voice-transcribe-handles-real-time-speech.txt", "jsonld": "https://wpnews.pro/news/metas-muse-voice-transcribe-handles-real-time-speech.jsonld"}}