Meta (META.US) has entered the real-time speech transcription market: benchmark tests outperform OpenAI and Alphabet Inc. Class C (GOOGL.US), with API prices as low as $0.18 per hour.

Recently, Meta (META.US) launched its real-time speech recognition model, Muse Voice Transcribe, through its Superintelligence Labs. This model has exceeded competitors such as OpenAI and Alphabet Inc. Class C (GOOGL.US) in some benchmark tests and is entering the market with an API price of $0.18 per hour. However, Meta has confirmed that the model will not have publicly available weights. According to reports, Muse Voice Transcribe achieved a word error rate of 3.1% on the AA-WER Streaming English speech transcription accuracy benchmark by Artificial Analysis, outperforming Cartesia Ink-2's 3.4%, ElevenLabs Scribe v2 Real-times 3.6%, OpenAI GPT Live Transcribes 3.9%, and Alphabet Inc. Class C Gemini 3.5 Transcribe Lives 4.0%. In terms of speaker differentiation ability, all models overall performance still leaves users unsatisfied, but Muse Voice Transcribe has an average error rate of 17.5% across multiple standard benchmarks, also leading the field. Meta reports that this model can distinguish over 20 speakers and supports more than 70 languages, with 25 being "widely validated." It can also manage language-switching among multiple speakers during conversation and can handle long conversations exceeding one hour. Muse Voice Transcribe is currently available through the Meta Model API, Meta AI for Mac, and Muse Code. The API pricing is $3 per 1,000 audio minutes, equivalent to $0.18 per hour. Unlike Meta's earlier Muse Glimmer series, which had open weights, a Meta spokesperson confirmed to The New Stack that this model's weights will not be made public. On a technical level, Muse Voice Transcribe belongs to the autoregressive multimodal model family, Muse Spark. Audio is input in 80-millisecond units, with approximately 12.5 blocks per second, each compressed into a single soft token. The model chooses on each block: either to output a text token or to output a special "next audio" placeholder, waiting for more audio context. When the audio ends, an "empty audio" token triggers the model to output the remaining text. This mechanism allows the model to control latency autonomously, referred to by Meta as "adaptive latency." Simple vocabulary can be transcribed almost instantly, while complex vocabulary benefits from additional audio context. This trade-off is trained during the reinforcement learning phase, optimizing the word error rate reward and delay reward through multiplication rather than addition. Speaker identification uses a similar mechanism. This summer, real-time speech transcription has become one of the most fiercely competitive sub-markets in the AI field. Recently, companies like OpenAI, Alphabet Inc. Class C, xAI, and Alibaba Group Holding Limited Sponsored ADR (BABA.US) have intensively released streaming speech models within weeks, while specialized vendors have been working in this area for some time. Muse Voice Transcribe's current benchmark lead of 0.3 percentage points is hard to maintain in such competitive intensity. However, for Meta, core products like smart glasses and Mac applications require real-time speech capabilities as fundamental support. This intrinsic demand may drive Meta to continue investing in this arena rather than merely pursuing external API market share.