Gemini 3.5 Transcribe: Google's New Speech Model

Google launched Gemini 3.5 Transcribe on 26 August 2026, cleaning up filler words and speaker self-corrections automatically across more than 85 languages.

Google introduced Gemini 3.5 Transcribe on 26 August 2026, a speech-to-text model built around the idea that raw transcription is rarely what anyone actually wants. Rather than reproducing every word as spoken, the model removes filler words such as "ums" and "ahs", automatically formats the output, and resolves speaker self-corrections — Google's own example being "let's meet Tuesday—no, Wednesday", which the model is designed to render as the corrected intent rather than the literal utterance [1].

Google says the model automatically detects and transcribes more than 85 languages, supports custom vocabulary for specialist jargon and unusual spellings, and identifies up to three speakers with timestamps, with support beyond three described as experimental [1]. It is available now in public preview through the Gemini API via Google AI Studio, Google Antigravity and the Gemini Enterprise Agent Platform, and reaches consumers through the Rambler feature in Gboard on Android and the Gemini app on macOS, with Chrome listed as coming soon [1].

The accuracy figures, and where they come from

Google cites word error rate (WER) measurements it attributes to Artificial Analysis, an independent benchmarking organisation: 4.0% WER for streaming transcription and 2.6% for non-streaming [1]. On the FLEURS multilingual benchmark, Google reports 5.50% WER streaming and 5.04% non-streaming [1]. The company also claims a 70% improvement in time-to-final-transcription compared with Chirp 3, its previous speech model [1].

This is a slightly stronger evidential position than most model announcements we have covered this month, because the headline accuracy figures are attributed to a third-party evaluator rather than to Google's own internal testing. The caveat is that Google has selected which third-party figures to publish and in which framing, and Next AI Compare has not independently reproduced them. The 70% latency improvement, notably, is a comparison against Google's own previous model rather than against any competitor.

Why "structured" matters more than "accurate"

Speech-to-text accuracy has been good enough for several years that WER alone is no longer the deciding factor for most business use. The remaining friction is that an accurate transcript of a person talking is still an awkward document: full of restarts, hesitations, corrections and no paragraph structure. Someone still has to clean it up before it becomes meeting notes, a summary or a draft.

Gemini 3.5 Transcribe targets that cleanup step rather than the transcription step. Whether that is genuinely useful depends on the use case, and the trade-off runs in both directions — a model that silently discards a self-correction is making an editorial judgement about intent, which is helpful for meeting notes and potentially problematic for a legal, medical or journalistic record where the literal utterance is what matters. Google's announcement does not indicate whether the cleanup behaviour can be disabled to obtain a verbatim transcript.

An unusual addition: function calling inside a transcription model

The most architecturally interesting claim is that the model can delegate tasks to other Gemini models mid-transcription, with Google citing image generation and file analysis as examples [1]. This blurs the line between a transcription model and an assistant: a spoken instruction could, in principle, trigger an action rather than simply being written down. Google has not detailed how the model distinguishes an instruction from speech that merely sounds like one, which is the obvious failure mode for such a feature.

How this sits against the alternatives

The transcription market is not short of credible options, and this release does not obviously displace them. ElevenLabs has been actively developing its Scribe line, adding realtime entity detection to Scribe v2 in a changelog entry dated 3 August 2026, and shipped a Dubbing v2 API on 10 August 2026 covering more than 90 languages while preserving speaker voice and pacing [2]. Deepgram remains a widely used API-first choice for developers building realtime speech recognition. OpenAI maintains its own transcription models within its audio lineup.

Cross-vendor WER comparisons are genuinely difficult to make responsibly, because results depend heavily on the test dataset, audio conditions and whether streaming or batch processing is measured — which is why we are not reproducing competitor figures alongside Google's here. What can be said is that Google's differentiation in this release is less about raw accuracy than about post-processing and distribution: it is shipping directly into Gboard and Chrome, which is a reach advantage no independent transcription vendor can match.

Why this matters

Transcription is becoming an input layer for other AI work rather than a standalone product. Meeting recordings feed summarisation, voice notes feed drafting, and support calls feed CRM records. A model that outputs clean, formatted, speaker-attributed text reduces the work needed before that downstream step. Google shipping it directly into Gboard and Chrome rather than only as an API also means a large number of people will use it without ever choosing it.

Who should care

Anyone dictating regularly on Android already has access through Gboard's Rambler feature and will encounter the cleanup behaviour by default. Developers building meeting-notes, voice-note or call-analysis products should evaluate the API preview, particularly the custom vocabulary support if their domain involves specialist terminology. Teams handling multilingual audio should note the 85-language automatic detection, which removes the need to specify a language in advance. Anyone whose use case requires a verbatim record — legal, compliance, medical, journalistic — should establish whether the cleanup can be turned off before adopting it.

Practical implications for buyers and users

The three-speaker limit is the most consequential practical constraint for business use, since a great many meetings involve four or more participants and Google describes support beyond three as experimental. Test with your actual meeting sizes rather than a two-person sample. Because everything on the developer side is in public preview, terms, behaviour and availability may change before general release, so this is a stage for evaluation rather than production commitment. The absence of published pricing in the announcement means cost per hour of audio cannot yet be compared against ElevenLabs, Deepgram or OpenAI, which is the comparison most buyers will actually need.

Limitations, availability and unresolved questions

Google has not published pricing for Gemini 3.5 Transcribe in the announcement. All three developer access routes are public preview rather than general availability, and no general-availability date is given. Chrome support is stated as forthcoming without a date. There is no published guidance on whether verbatim, uncleaned output is available as an option, nor on how audio data is retained or used — a material question for anyone transcribing confidential meetings. Google has also not detailed accuracy variation across the 85-plus supported languages, and aggregate FLEURS figures can conceal significant per-language differences.

Verdict

Gemini 3.5 Transcribe is a well-targeted release that competes on output usability rather than headline accuracy, and the decision to cite third-party evaluation figures rather than purely internal ones is more credible than the industry norm. Its distribution through Gboard and Chrome is arguably a bigger competitive advantage than the model itself. The three-speaker ceiling limits its immediate usefulness for meeting transcription specifically, and the open question of whether verbatim output is available should be resolved before anyone adopts it for a record-keeping use case. Worth testing now in preview; too early to standardise a workflow on.

The current AI Voice & Audio shortlist

Where this sits in the wider market: our current shortlist for AI Voice & Audio, what each tool is best at and the main caution to check before committing.

Related reading

• AI Voice & Audio — Next AI Compare's AI voice and audio hub, covering ElevenLabs, Deepgram and Gemini Audio
• Best AI Voice and Audio Tools in 2026: Production Comparison — the voice and audio comparison this model belongs in
• 7 Best AI Voice Generators in 2026 — shortlist guide covering transcription alongside synthesis
• A Human-Reviewed AI Content Workflow That Avoids Generic Output — relevant to the verbatim-versus-cleaned transcript question
• Adobe Firefly Adds Music, Speech and Sound Effects
• ElevenLabs vs PlayHT vs Murf: Which Voice Platform Wins in 2026?

Sources and verification notes

Primary product documentation checked for this update: