Blab AI
Meta Launches Muse Voice Transcribe, a Real-Time Speech-to-Text Model That Tops Current Benchmarks
Meta Platforms Inc. revealed Muse Voice Transcribe on Tuesday, its first real‑time audio perception model. The new offering, part of the Muse Spark suite from Meta Superintelligence Labs, streams speech‑to‑text, speaker diarization and endpointing within a single autoregressive framework. Meta says the model can handle more than 20 simultaneous speakers, supports over 70 languages—25 of which have undergone extensive verification—and can transcribe conversations that stretch beyond an hour. The model is accessible via the Meta Model API, Meta AI for Mac, and Muse Code, priced at $3.00 per 1,000 audio minutes (about $0.18 per hour). A spokesperson confirmed to The New Stack that Meta will not release the model’s weights.
Benchmark tests place Muse Voice Transcribe at the top of the Artificial Analysis (AA) Streaming speech‑to‑text accuracy leaderboard. With a word error rate of 3.1%, it eclipses rivals such as Cartesia Ink‑2 (3.4%), ElevenLabs’ Scribe v2 Real‑time (3.6%), GPT Live Transcribe (3.9%) and Gemini 3.5 Transcribe Live (4.0%). In speaker diarization, the model records a 17.5 % error rate across several standard benchmarks, again outpacing the same set of competitors. Those figures reflect English‑only evaluations.
At its core, Muse Voice Transcribe is an autoregressive, multimodal architecture that ingests audio in 80‑millisecond slices, compressing each into a single soft token. For every chunk, the model decides whether to emit a text token or a special “next audio” placeholder, which the system later replaces with the following segment. When the stream ends, an “empty audio” token prompts the model to flush any buffered text. This design gives the system direct control over latency—a capability Meta calls “adaptive delay.” During training, the reinforcement‑learning phase multiplies the word‑error‑rate reward by the delay reward, rather than adding them, encouraging the model to balance precision and speed. A comparable mechanism governs speaker detection.
Meta’s entry comes at a time when the real‑time transcription market is crowded. OpenAI, Google, xAI and Alibaba have all unveiled streaming models in recent months, while specialist providers have been active for longer. Analysts note that a 0.3‑point lead on a benchmark is unlikely to last, yet it positions Meta as a serious contender. The company’s broader ecosystem—system‑wide dictation on macOS, the Meta AI desktop app, and its upcoming glasses—relies on robust real‑time speech recognition, giving Meta a built‑in incentive to refine the technology.
Introduced in April 2026, Meta’s Muse Spark family has focused on multimodal reasoning, coding assistance and tool use. Muse Voice Transcribe extends that line into audio perception, marking the first time Meta has released a dedicated real‑time ASR model. The decision not to open the weights aligns with Meta’s policy on other proprietary models such as Muse Glimmer.
At present, the model is only available through paid API access and the Meta AI for Mac application. No public release of the weights or a free tier has been announced. The pricing strategy—$3 per 1,000 minutes—positions the service competitively against Google Cloud’s speech‑to‑text pricing, which is roughly 80 % higher for comparable usage.
In summary, Meta has launched a new real‑time speech‑to‑text model that, according to current benchmarks, outperforms existing commercial offerings in English. The model supports multilingual transcription, speaker diarization, and long‑form conversations, and is integrated into Meta’s API, desktop, and developer tools. Future efforts are likely to focus on expanding language coverage, tightening speaker accuracy, and monitoring performance in the rapidly evolving real‑time transcription market.
Benchmark tests place Muse Voice Transcribe at the top of the Artificial Analysis (AA) Streaming speech‑to‑text accuracy leaderboard. With a word error rate of 3.1%, it eclipses rivals such as Cartesia Ink‑2 (3.4%), ElevenLabs’ Scribe v2 Real‑time (3.6%), GPT Live Transcribe (3.9%) and Gemini 3.5 Transcribe Live (4.0%). In speaker diarization, the model records a 17.5 % error rate across several standard benchmarks, again outpacing the same set of competitors. Those figures reflect English‑only evaluations.
At its core, Muse Voice Transcribe is an autoregressive, multimodal architecture that ingests audio in 80‑millisecond slices, compressing each into a single soft token. For every chunk, the model decides whether to emit a text token or a special “next audio” placeholder, which the system later replaces with the following segment. When the stream ends, an “empty audio” token prompts the model to flush any buffered text. This design gives the system direct control over latency—a capability Meta calls “adaptive delay.” During training, the reinforcement‑learning phase multiplies the word‑error‑rate reward by the delay reward, rather than adding them, encouraging the model to balance precision and speed. A comparable mechanism governs speaker detection.
Meta’s entry comes at a time when the real‑time transcription market is crowded. OpenAI, Google, xAI and Alibaba have all unveiled streaming models in recent months, while specialist providers have been active for longer. Analysts note that a 0.3‑point lead on a benchmark is unlikely to last, yet it positions Meta as a serious contender. The company’s broader ecosystem—system‑wide dictation on macOS, the Meta AI desktop app, and its upcoming glasses—relies on robust real‑time speech recognition, giving Meta a built‑in incentive to refine the technology.
Introduced in April 2026, Meta’s Muse Spark family has focused on multimodal reasoning, coding assistance and tool use. Muse Voice Transcribe extends that line into audio perception, marking the first time Meta has released a dedicated real‑time ASR model. The decision not to open the weights aligns with Meta’s policy on other proprietary models such as Muse Glimmer.
At present, the model is only available through paid API access and the Meta AI for Mac application. No public release of the weights or a free tier has been announced. The pricing strategy—$3 per 1,000 minutes—positions the service competitively against Google Cloud’s speech‑to‑text pricing, which is roughly 80 % higher for comparable usage.
In summary, Meta has launched a new real‑time speech‑to‑text model that, according to current benchmarks, outperforms existing commercial offerings in English. The model supports multilingual transcription, speaker diarization, and long‑form conversations, and is integrated into Meta’s API, desktop, and developer tools. Future efforts are likely to focus on expanding language coverage, tightening speaker accuracy, and monitoring performance in the rapidly evolving real‑time transcription market.
