Overview
Gemini 3.5 Transcribe is Google's August 26, 2026 release of two dedicated speech-to-text models: gemini-3.5-transcribe for non-streaming audio-file transcription and gemini-3.5-transcribe-live for real-time streams. Both are generally available in the Gemini API. They are one product release with different transport and latency profiles, so teams can use the same model generation for uploaded recordings and live captions without treating them as separate tools.
This is not a Gemini chatbot update. The models return transcripts rather than conversational answers, and they succeed Google's Chirp speech-recognition line. The practical upgrade is a lower-latency, more accurate recognition layer with 85+ languages, code-switching, vocabulary prompting, and optional transcript structure.
What's New
Dedicated non-streaming and live models
The non-streaming endpoint runs through the Interactions API and processes recorded audio with synchronous unary requests. The live endpoint uses the Live API over WebSockets, returning partial and final results as audio arrives. This split lets developers optimize for full-recording accuracy or streaming latency while staying within the Gemini API.
Multilingual and domain controls
Gemini 3.5 Transcribe detects more than 85 languages and handles code-switching. Developers can provide vocabulary lists with names, acronyms, jargon, or product terminology; the API accepts up to 1,000 entries, although Google says 100 or fewer usually produces the best result.
Structured or cleaned transcripts
Non-streaming transcription supports speaker diarization and word-level timestamps for meetings, interviews, captioning, and media search. Smart transcription takes the opposite approach: it removes filler words, resolves spoken corrections, and formats the text for readability. Smart transcription cannot be combined with diarization or word timestamps in the same request.
Performance Benchmarks
Google's launch post cites Artificial Analysis for the 4.0% streaming WER, 2.6% non-streaming WER, and 70% time-to-final improvement; Google separately reports the FLEURS results. These benchmarks should still be verified on each team's languages, acoustics, and vocabulary.
| Measurement | Streaming | Non-streaming | Selection value |
|---|---|---|---|
| Word error rate | 4.0% | 2.6% | Lower is better; non-streaming has more context before final output |
| FLEURS word error rate | 5.50% | 5.04% | Multilingual benchmark across diverse languages |
| Time to final transcription vs Chirp 3 | 70% faster | — | Most relevant to live captions and voice interfaces |
The comparison is strongest as a directional upgrade signal from Chirp 3. Production teams should evaluate named entities, accented speech, overlapping speakers, poor microphones, and domain vocabulary rather than relying on an aggregate WER alone.
Compared With Previous Version
Chirp 3 is the direct predecessor named in Google's launch material. Gemini 3.5 Transcribe introduces dedicated non-streaming and live model IDs in the Gemini API alongside the separately available Chirp 3 Cloud Speech-to-Text V2 model, while improving both recognition and final-result latency in Google's tests.
| Area | Chirp 3 | Gemini 3.5 Transcribe |
|---|---|---|
| Primary developer surface | Google Cloud Speech-to-Text V2 | Gemini API, plus separate Google Cloud enterprise endpoints |
| Non-streaming and streaming | Available through Cloud speech workflows | Dedicated non-streaming and live Gemini model IDs |
| Language behavior | Universal multilingual speech recognition | 85+ languages, automatic detection, and code-switching |
| Prompt controls | Cloud Speech adaptation features | Custom vocabulary plus smart-transcription instructions |
| Launch comparison | Baseline | Google reports lower WER and a 70% improvement in time to final transcription |
The move is not necessarily a drop-in model-ID replacement. Chirp users should compare request schemas, regional controls, data-governance requirements, quotas, and endpoint maturity before migrating.
Compatibility Notes
- Non-streaming model ID:
gemini-3.5-transcribe, accessed through the Interactions API. - Live model ID:
gemini-3.5-transcribe-live, accessed through the Live API over WebSockets. - Audio duration: Non-streaming unary requests support up to one hour, or 30 minutes when speaker diarization or word-level timestamps are enabled. Live transcription sessions support continuous streaming for up to 10 minutes.
- Feature combinations: Smart transcription cannot be combined with diarization or word-level timestamps.
- Speaker attribution: Google documents up to eight speakers for diarization, while attribution with three or more speakers remains experimental.
- Channel stage: The August 26 Gemini API changelog lists the stable Gemini API IDs as GA, while Gemini Enterprise Agent Platform documents
-previewIDs; Google's August 26 launch blog still says public preview, so enterprise teams should verify the exact endpoint and service-level terms for their channel.
Store raw audio and expected transcripts for regression testing. Streaming applications should also handle session rotation, reconnection, partial-result replacement, and late finalization without duplicating text.
Pricing & Plans
| Mode | Audio input | Text output | Approximate combined rate |
|---|---|---|---|
| Non-streaming | $2.00 / 1M tokens, about $0.003/min | $12.00 / 1M tokens, about $0.002/min | About $0.005/min |
| Live | $3.50 / 1M tokens, about $0.005/min | $21.00 / 1M tokens, about $0.004/min | About $0.009/min |
A free tier is available for evaluation. Actual cost depends on audio duration and output-token volume, while Google Cloud enterprise contracts, quotas, and regional deployment requirements may differ from public Gemini API pricing.
Who Should Upgrade / Who Should Wait
Upgrade or start testing now if:
- You use Chirp 3 and need lower time-to-final latency for live captions or voice input.
- Your recordings switch languages or contain domain terms that benefit from custom vocabulary.
- You want non-streaming and live transcription in the same Gemini developer platform.
- You need speaker labels, word timestamps, or cleaned readable transcripts and can choose among those modes per request.
Wait or run a longer evaluation if:
- You require a specific Google Cloud region, enterprise SLA, or stable non-preview endpoint outside the Gemini API.
- Your workflow depends on long recordings with diarization or timestamps beyond the 30-minute request limit.
- You need an end-user meeting workspace, summaries, action items, telephony, or voice-agent orchestration rather than a speech-to-text API.
- Your audio has heavy overlap, many speakers, or a regulated accuracy threshold that has not yet passed an internal evaluation.
