Gemini 3.5 Transcribe: 2.6% WER, 85+ Languages, Two API Surfaces

Google shipped Gemini 3.5 Transcribe on August 27, 2026, as a managed speech-to-text service built around two distinct API surfaces. Artificial Analysis benchmarks place the non-streaming variant at 2.6% average word error rate and the streaming variant at 4.0%, with time to final transcription dropping 70% relative to Chirp 3, the model it supersedes. For engineers evaluating ASR pipelines, the architectural split between batch and real-time paths — each with its own feature set, rate limits, and pricing — is the decision that requires the most upfront planning.

Two Endpoints With Incompatible Feature Sets

The gemini-3.5-transcribe model operates through the Interactions API and targets pre-recorded audio files. It supports speaker diarization with up to 8 speaker labels, word-level start and end offsets, and custom vocabulary biasing accepting up to 1,000 terms (with documented best results below 100 terms). Standard requests accept up to one hour of audio; enabling diarization or word-level timestamps reduces that ceiling to 30 minutes. Attribution beyond 2 speakers is flagged as experimental.

The gemini-3.5-transcribe-live model operates through the Live API over a bidirectional WebSocket. It emits interim_input_transcription events — speculative partials during active speech — followed by authoritative input_transcription events when a turn finalizes. Audio ingestion format is 16-bit PCM at 16 kHz mono delivered in 100ms chunks. Voice activity detection runs in automatic, hybrid, or manual modes. Ephemeral tokens allow browser and mobile clients to stream directly without exposing a persistent API key. Hard constraints: sessions cap at 10 continuous minutes, and neither speaker diarization nor word-level timestamps are available on this surface.

This is a capabilities fork, not merely a latency tradeoff. Builders who need an auditable transcript with speaker attribution must use the Interactions API. Those building sub-second voice agents must use the Live API and accept its limitations. The structural logic echoes a pattern covered in pipeline architecture driving AI gains in 2026: how components connect matters as much as underlying model quality.

Transcription Modes and the Smart/Verbatim Tradeoff

Both endpoints expose two output modes. verbatim is the default, preserving all fillers, repetitions, and false starts. smart mode removes disfluencies, resolves inline self-corrections, and applies structured formatting — Google's documented example converts "Um, so for the meeting, I think we should, uh, invite Alice and, wait no, Bob and Carol" into "For the meeting, I think we should invite Bob and Carol." Smart mode also converts sequential spoken enumeration into formatted list output.

The critical constraint: smart mode cannot be combined with word-level timestamps or speaker diarization. A readable, cleaned output and a timestamped, speaker-attributed transcript require separate calls.

Benchmark Performance and Language Coverage

Language detection is automatic across more than 85 locales, with mid-sentence code-switching handled without developer configuration. No per-language WER breakdown was published alongside the aggregate figures.

Deployment Model and Pricing

There are no open weights and no self-hosted path. Gemini 3.5 Transcribe is exclusively a managed API service. Google structures access across three tiers: solo developers and startups use the free tier via Google AI Studio; the paid tier raises rate limits and guarantees audio content is not used for product improvement; regulated enterprises route through the Gemini Enterprise Agent Platform, which adds provisioned throughput, compliance controls, and volume discounts. Both developer and enterprise tracks are in public preview.

Blended cost runs approximately $0.005 per minute for batch and $0.009 per minute for the live endpoint, with billing based on token consumption — 25 audio tokens per second inbound and 175 text tokens per minute outbound — meaning actual cost varies with speech density. On the integration side, the Live API is already wired into LiveKit, Pipecat, Agora, Fishjam, Vercel, and Vision Agents. Consumer deployments include Rambler on Android, the Gemini app on macOS, and Google Antigravity, with Chrome listed as forthcoming.

The managed-only decision limits adoption for teams with data residency requirements or self-hosting cost structures, but it also means Google controls the full inference stack. At 2.6% non-streaming WER across 85+ languages with a 70% latency improvement over Chirp 3, the baseline capability is credible; the question for most engineering teams is whether the dual-endpoint constraints fit their architecture before the public preview hardens into a production dependency.