Google, Meta, and Microsoft ship new speech-to-text models for agents

Google, Meta, and Microsoft each published speech-to-text models for voice agents and live workflows — software that turns talk into text — and the word-error and speed scores stay the companies’.

Three big labs shipped voice-to-text models in about a week, and the posted prices are now on the record. Word-error and speed leaderboard claims stay the companies’.

Within about a week, Google, Meta, and Microsoft each published dedicated speech-to-text (STT) models — software that turns spoken audio into written words — aimed at voice agents and live workflows. DeepLearning.AI’s The Batch dated that cluster 11 September 2026. That recap is the filing event for the three-lab week; each lab’s own page is the ship.

Google’s 26 August 2026 post introduced Gemini 3.5 Transcribe — “our most precise speech-to-text model yet, designed for intelligent voice interactions” — in Google AI Studio and Gemini Enterprise Agent Platform. Google lists 85+ languages; smart transcription (self-corrections, filler cleanup, auto-format); custom vocabulary; and word-level timestamps on files. Live-stream and file model IDs are in Sources.

On who spoke: Google’s blog says pre-recorded speaker labels work “for up to three speakers (support for 3+ speakers is experimental).” API docs also list file diarization — labeling who spoke — “up to 8 speakers,” with 3+ still experimental, and say live streaming has no diarization. Google cites Artificial Analysis word error rate (WER, the share of words wrong; lower is better) of about 4.0% streaming and 2.6% on files. Those benches are Google’s citation, not a desk measurement.

Meta Superintelligence Labs published Muse Voice Transcribe on 1 September 2026 — “the first real-time audio perception model” from that lab. Meta claims live speech recognition plus diarization (20+ speakers claimed) plus endpointing, which is deciding when a speaker has finished. Adaptive delay: 80 ms audio chunks — 80 thousandths of a second — and the model chooses when to emit text. Training covers 70+ languages, with 25 “extensively verified.” Long audio “exceeding one hour” needs no required post-processing, Meta says.

The same Meta page: one-click voice dictation in Meta AI and Muse Code, “hold ‘Fn’ to try.” Meta’s API docs price it at $0.18 per hour of audio processed — live and files the same — which is $3 per 1,000 minutes. Meta also claims first on Artificial Analysis live speech-to-text and on public diarization benches as of 1 September 2026 — Meta’s ranking.

Microsoft AI published MAI-Transcribe-2 on 3 September 2026. Microsoft claims it is “the fastest, most accurate and cheapest speech recognition model in the world” versus Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large, and Scribe V2. Posted launch price: $0.10 per hour of audio “as a limited-time offer until the end of the year.” Microsoft lists 60 languages, speaker diarization, word-level timestamps, keyword biasing, and verbatim vs clean styles.

Microsoft cites FLEURS 5.2% WER (first across 60 languages on that bench, it says) and an Artificial Analysis accuracy-versus-speed lead / second on the WER board. Its page: Artificial Analysis evals show the model 10× faster than GPT-Transcribe and 5× faster than Gemini 3.5 Transcribe. The Batch says Microsoft states an hour of audio can be transcribed in about 10 seconds. Those speed and WER lines stay Microsoft’s.

Availability on the primaries: Google via Gemini API / AI Studio / Enterprise Agent Platform. Meta via Meta AI (Mac dictation), Muse Code, and Meta Model API. Microsoft’s news page: demo through Microsoft Foundry, MAI Playground, and Open Router. Microsoft Learn: MAI-Transcribe-2 is in public preview on Azure Speech / Foundry Tools — “without a service-level agreement, and is not recommended for production workloads.”