Meta's Muse Voice Transcribe Runs at $0.18 an Hour

Seventy languages and twenty-five languages. For Muse Voice Transcribe, the speech recognition model Meta introduced on September 1, the gap between those two numbers says more than the rest of the announcement. The model was trained on 70+ languages, but the count Meta describes as extensively verified is 25, and the company recommends sticking to those.

What the numbers claim

The headline claim is about speed. Meta reports final transcription in 0.16 seconds using an adaptive delay technique, while holding word error rate at 3.0%. For live captioning or note-taking during a call, that latency figure is the one that decides whether a product feels usable.

The second claim is about diarization, the job of telling speakers apart. The model handles 20+ speakers with a 17.5% diarization error rate on benchmark datasets, and Meta says that as of September 1 it ranks first on Artificial Analysis for streaming speech-to-text and on public diarization benchmarks.

Worth translating that error rate into practice: in a six-person meeting recording, roughly one part in six of the speech could be attributed to the wrong person. Anywhere the question who said this actually matters, such as interviews, board minutes or a dispute file, a human still has to review the output.

The technology reaches consumers too. Meta is shipping the same engine as real-time dictation that works inside any Mac application, so the desktop feature and the enterprise API share a foundation.

Pricing does not appear in Meta's own announcement. The figure of $3 per 1,000 audio minutes through the Model API, about $0.18 an hour, comes from secondary reporting, which also says speaker labeling is included rather than billed separately.

Our read

At $0.18 an hour, the economics change at call-center scale. A team processing 1,000 hours of speech a month lands near $180. A year ago that same volume cost several times as much with most providers. Meeting summaries, call quality analysis and interview archives no longer have a budget problem, they have an accuracy problem.

Which brings the language gap back into focus. If the language you operate in is not on Meta's verified list, the 3.0% error rate is not a number you can plan around. The model may well handle your language, it simply has not been presented as validated for it, and those are different claims.

The pattern we keep seeing in the field is that speech models do fine on general conversation and degrade quickly as domain jargon and proper nouns pile up. An insurance call about policy clauses and an order taken at a coffee counter are not the same difficulty, even in the same language.

So evaluate this on your own recordings rather than on the leaderboard. Take 30 minutes of real audio pulled at random from your archive, run it through both your current provider and Muse, and compare the transcripts word by word. That half day of work answers the licensing question better than any benchmark table will.

Sources: Meta AI Research, Meta Developer, TechRepublic