Gemini 3.5 Transcribe splits real-time and batch speech-to-text, Google says
Google DeepMind
- Models
- Developer tools
- #gemini
- #speech-to-text
- #chirp
Google introduced Gemini 3.5 Transcribe, a speech-to-text model it says cleans disfluencies, applies custom vocabulary and detects 85+ languages. It arrives as two APIs: a streaming model for low-latency voice apps and a pre-recorded model with speaker attribution and timestamps. Google says it beat its prior Chirp 3 model on Artificial Analysis word error rates and reported a 70% latency improvement.
What Google announced
Google introduced Gemini 3.5 Transcribe, which it calls its most precise speech-to-text model, positioned against recognizers that struggle with noise, jargon and disfluent speech. According to the announcement, the model converts raw audio into formatted text, handles self-corrections, strips filler words and adapts to a supplied custom vocabulary. It is offered as two interfaces: a streaming mode for continuous, bidirectional audio with sub-second latency, and a pre-recorded mode that returns speaker attribution and word-level timestamps. Google says the model auto-detects over 85 languages and attributes speakers for up to three voices, with more described as experimental. The accuracy numbers cited come from Artificial Analysis (4.0% WER streaming, 2.6% non-streaming) and FLEURS (5.50% and 5.04%), and Google reports a 70% improvement in time to final transcription over Chirp 3.
Where a support team might try it
A customer-support operations lead could consider the pre-recorded interface for post-call quality review, where the stated word-level timestamps and up-to-three-speaker attribution would let reviewers jump to specific moments and separate agent from caller speech. The custom vocabulary feature is the natural experiment target: a team could supply its product names, order-ID formats and postal codes, then compare raw transcripts against the same recordings processed without that vocabulary list. This is a hypothesis to test, not a demonstrated outcome; the excerpt reports strong entity handling but gives no per-domain results.
A small trial and how to judge it
A team could start with a held-out set of its own recordings rather than published benchmarks. Steps: (1) collect a baseline by transcribing the set with the current pipeline and recording word error rate, entity accuracy on IDs and codes, and time to final transcript; (2) repeat with gemini-3.5-transcribe, first without and then with a curated custom vocabulary; (3) for any streaming use case, measure perceived latency on the live path separately. Success means lower entity errors without higher overall WER. Failure checks: manually spot-check fluent-sounding but wrong words, since auto-formatting can mask errors, and audit attribution accuracy on three-speaker calls. Human review of a sample should stay in place before transcripts drive any decisions.
Trade-offs and when to hold back
The two-interface split means teams must choose per use case: streaming for interactivity, batch for richer output such as speaker labels. Smart cleanup is a trade-off, since removing filler words and correcting self-revisions changes the record, which may not suit verbatim legal or clinical transcripts. Prerequisites include access through the Gemini API, AI Studio or the enterprise platform, but the excerpt gives no pricing, quotas or data-handling terms, so cost and compliance need separate confirmation. Pushback from the announcement's own framing: figures are vendor-cited, support beyond three speakers is experimental, and Chrome availability is only 'coming soon'. Teams needing audited accuracy on four-plus speakers should wait or test carefully.
Context and limitations
- Assessment is based on a short excerpt of Google's announcement; the full post was not analyzed.
- Accuracy and latency figures are Google's citations of Artificial Analysis and FLEURS, not independent testing; no pricing, licensing or SLA details appear in the excerpt.
- Speaker attribution is stated as accurate for up to three speakers; support beyond three is described as experimental.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.