Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is a family of speech recognition models developed by Google DeepMind and introduced by Google on August 26, 2026. It converts spoken audio to text through separate recorded-audio and live-streaming variants. Google launched both variants in public preview through the Gemini Developer API and Gemini Enterprise Agent Platform, with different model identifiers on the two platforms.^{[1][3][8][10]}

The family is specialized for transcription. The recorded endpoint can return speaker labels and word-level timestamps, while the live endpoint returns interim and finalized text as audio arrives. Both support automatic language detection, custom vocabulary biasing, and a smart mode that edits disfluent speech into more readable text. Smart output is not a verbatim record because it can remove filler words, resolve spoken corrections, and change formatting.^[3][4][5]

Model overview

Google DeepMind's model card groups Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live within the Gemini 3.5 Audio family. It says both are based on Gemini 3 Pro, accept audio and text within a context window of up to 96,000 tokens, and produce text with an output limit of 32,000 tokens. The card refers to the Gemini 3 Pro documentation for architecture, training data, and training-data processing rather than disclosing a separate transcription-specific architecture or corpus.^[10]

The model card gives the family a January 2025 knowledge cutoff. It also warns that the transcription variants can exhibit foundation-model limitations such as hallucinations, along with occasional slowness or timeouts. The API documentation describes the model as a speech-to-text system, not a general audio question-answering model or a text-to-speech system.^[4][10]

Scroll sideways for more →

Google's launch materials also described integrations in Gboard, the Gemini app, Google Antigravity, and other products. Those applications can add their own context and orchestration. In particular, the launch blog limited its function-calling example to the Gemini app on macOS, where other Gemini models handle background tasks. The Developer API model page lists function calling as unsupported for gemini-3.5-transcribe, so the app behavior is not a documented capability of the transcription API itself.^[1][3]

Endpoints and model identifiers

The platform name determines both the model identifier and the request interface. The Gemini Developer API uses model IDs without a preview suffix. Gemini Enterprise Agent Platform uses preview-suffixed IDs and was available only through the global location at launch.^[3][8]

Scroll sideways for more →

The two official Agent Platform pages therefore did not agree on the recorded-audio limit at launch. The model page's feature table specified 15 minutes, while the pricing page described synchronous processing of files up to 60 minutes. This difference should not be collapsed into a single family-wide limit. An implementation must follow the documentation and validation behavior of the exact endpoint it calls.^[8][9]

For recorded audio in the Developer API, Google's primary guide uploads a file through the Files API and passes its URI to an Interactions request. The legacy Generate Content guide documents the same gemini-3.5-transcribe model through generateContent. These are alternative documented request surfaces for the recorded variant, not separate transcription models.^[4][6]

Transcription modes

Both recorded and live variants support two output modes. The default is called verbatim. Google defines it as retaining filler words, repetitions, pauses, and false starts. The smart mode applies post-processing for readability, including disfluency removal, resolution of inline self-corrections, punctuation and casing, inverse text normalization, and structured formatting such as paragraphs and lists.^[4][5]

Scroll sideways for more →

The name verbatim describes the requested behavior, not a guarantee of an error-free record. Google's own model card acknowledges that the model can hallucinate. Smart mode has a different goal: it can deliberately omit spoken material and replace a self-correction with the speaker's final choice. A smart transcript is therefore edited dictation and should not be presented as a word-for-word transcript.^[4][10]

Language handling and vocabulary biasing

The API documentation lists automatic detection across more than 85 languages or locales. Detection can change within a session when speakers switch languages, and a caller can supply one or more BCP-47 language codes when the expected language is known. The documentation describes both changes between utterances and code switching within an utterance.^[3][4][5]

Callers can also provide a custom vocabulary of up to 1,000 terms to bias recognition toward technical terms, acronyms, brand names, and proper nouns. Google says results are typically best with no more than 100 targeted terms. Vocabulary biasing changes the model's preference; the documentation does not guarantee that every supplied spelling will appear in the output.^[3][4][5]

Support for code switching is a capability statement rather than evidence of equal performance for every language pair. Research on code-switched ASR evaluation has found that orthography, transliteration, and normalization choices affect how automatic metrics correspond to human judgments. The supported-language list and the benchmark results should therefore be read separately.^[16]

Speaker diarization and timestamps

The recorded variant can assign recurring voices labels such as spk_1 and spk_2. This is speaker diarization, the task commonly summarized as determining who spoke when. A speaker label groups detected speech segments; it does not establish the real-world identity of the person speaking.^[4][15]

The current Developer API guide allows up to eight speakers, but calls attribution for three or more speakers experimental. Google's launch announcement described multi-speaker attribution for up to three speakers and also said support beyond three was experimental. Taken together, the sources distinguish the documented maximum from the range with nonexperimental support rather than establishing reliable eight-speaker accuracy.^[1][3][4]

Recorded transcription can return start and end offsets for each recognized word. Google warns that enabling word-level timestamps may reduce overall transcription accuracy. Diarization and timestamps can be requested together, but both require verbatim mode and reduce the Developer API file limit from 60 to 30 minutes.^[3][4]

The live variant does not support speaker diarization or word-level timestamps. It returns interim hypotheses while a person is speaking and a finalized transcription when the speech turn ends. Google describes these as utterance-level events, which are different from per-word time annotations.^[5]

Live streaming behavior

Gemini 3.5 Transcribe Live uses the Gemini Live API's bidirectional connection but operates as a dedicated text transcription pipeline rather than a conversational voice agent. Its response modality is text, and its input stream is audio. The live transcription guide does not expose the broader live agent surface for spoken responses, Google Search, system instructions, or function calling.^[5]

The Developer API guide specifies raw 16-bit PCM audio, with mono little-endian audio at 16 kHz as the documented example. It recommends sending about 100 milliseconds of audio at a time and uses the MIME type audio/pcm;rate=16000. The server sends rapidly changing interim hypotheses and separately marks finalized text.^[5]

Voice activity detection determines when speech begins and ends. Server-side automatic detection is the default. A hybrid design can keep server-side start detection while a client sends audio_stream_end as soon as its local detector finds the end of speech. Manual mode disables automatic detection and uses explicit activity-start and activity-end signals, which supports push-to-talk interfaces.^[5]

Performance and evaluation

Word error rate is a standard ASR metric calculated as substitutions plus deletions plus insertions, divided by the number of words in the reference transcript. Lower values indicate fewer errors within the same evaluation setup. The value still depends on the audio, reference transcript, normalization rules, prompting, chunking, and endpoint behavior.^[12][14]

Scroll sideways for more →

Artificial Analysis says it tests provider APIs directly. Its AA-WER composite weights the proprietary AA-AgentTalk set at 50%, VoxPopuli-Cleaned-AA at 25%, and Earnings22-Cleaned-AA at 25%. The current method uses English audio and may split longer Earnings22 recordings into shorter chunks when endpoint constraints require it. These conditions make the 2.6% and 4.0% results well-specified benchmark results, but they do not establish accuracy across every supported language or every deployment environment.^[12]

Artificial Analysis asks configurable systems to transcribe verbatim and output only spoken words in sequence. Smart mode intentionally changes that sequence by removing or rewriting some spoken material. AA-WER therefore evaluates the literal-transcription use case and does not directly measure whether smart mode produces useful edited dictation.^[4][12]

The original FLEURS paper describes a parallel speech dataset spanning 102 languages, with about 12 hours of supervision per language. Google's evaluation report says its comparison used 26 top locales. The 5.50% and 5.04% figures apply to that selected multilingual subset, not to all 102 FLEURS languages or all 85+ locales documented for the API.^[11][13]

Google also said that time to final transcription improved by 70% relative to Chirp 3, attributing the measurement to Artificial Analysis. The launch materials do not turn that relative result into a universal latency guarantee, and live latency can also depend on endpointing and voice activity detection.^[1][5]

Pricing

At launch, the Gemini Developer API charged separately for audio input tokens and text output tokens. Google based its per-minute estimates on 25 audio tokens per second and 175 text tokens per minute. Actual token use can differ from those assumptions.^[7]

Scroll sideways for more →

The Developer API pricing page listed a free tier for both transcription variants. It also stated that free-tier use could be used to improve Google's products, while paid-tier use would not be used for that purpose. That distinction concerns the service tier and its terms, not the model's recognition quality.^[7]

Limitations and interpretation

Gemini 3.5 Transcribe launched as a preview service. Google documented several concrete constraints: a 10-minute live-session cap, no live speaker diarization, no live word-level timestamps, experimental speaker attribution for three or more speakers in recorded audio, possible accuracy loss when word timestamps are enabled, and incompatibility between smart mode and the annotation features.^[3][4][5][8]

The Developer API model page also lists caching, code execution, file search, function calling, image generation, and thinking as unsupported. Batch, flex, and priority inference are not listed as consumption options for this model. These limits keep the API's documented role narrower than the abilities of a general Gemini language model.^[3]

No benchmark result eliminates transcription errors. The model card's warning about hallucinations is especially relevant when a record depends on exact wording. Speaker labels, timestamps, custom vocabulary, and verbatim mode can structure a transcript, but none of them proves that the recognized words match the source audio. Smart mode adds a separate fidelity tradeoff because some changes are intentional.^[4][10]

References

1. ^Diego Melendo Casado and Luke Leonhard, Google. "Introducing Gemini 3.5 Transcribe." August 26, 2026. blog.google/...gemini-3-5-transcribe
2. ^Google AI Studio. "Introducing Gemini 3.5 Transcribe." X post, August 26, 2026. x.com/...2092659361235587416
3. ^Google AI for Developers. "Gemini 3.5 Transcribe." Accessed August 27, 2026. ai.google.dev/...gemini-3.5-transcribe
4. ^Google AI for Developers. "Audio transcription." Accessed August 27, 2026. ai.google.dev/...transcribe
5. ^Google AI for Developers. "Live transcription with Gemini Live API." Accessed August 27, 2026. ai.google.dev/...live-transcribe
6. ^Google AI for Developers. "Audio transcription: Gemini Generate Content API (Legacy)." Accessed August 27, 2026. ai.google.dev/...transcribe
7. ^Google AI for Developers. "Gemini Developer API pricing." Accessed August 27, 2026. ai.google.dev/...pricing
8. ^Google Cloud. "Gemini 3.5 Transcribe: Gemini Enterprise Agent Platform." Accessed August 27, 2026. docs.cloud.google.com/...3-5-transcribe
9. ^Google Cloud. "Agent Platform Pricing." Accessed August 27, 2026. cloud.google.com/...pricing
10. ^Google DeepMind. "Gemini 3.5 Audio (Live Translate, Transcribe, Transcribe Live): Model Card." August 26, 2026. deepmind.google/...gemini-3-5-audio
11. ^Google DeepMind. "Gemini 3.5 Audio (Transcribe, Transcribe Live): Model evaluation, approach, methodology and results." August 2026. deepmind.google/...gemini-3-5-transcribe
12. ^Artificial Analysis. "Speech to Text Benchmarking Methodology." Accessed August 27, 2026. artificialanalysis.ai/...methodology
13. ^Alexis Conneau et al. "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech." arXiv:2205.12446, 2022. arxiv.org/...2205.12446
14. ^National Institute of Standards and Technology. "NIST 2017 Pilot OpenSAT Evaluation Plan." March 5, 2018. nist.gov/...nist2017pilotopensatevalplanfinalpdf
15. ^Tae Jin Park et al. "A Review of Speaker Diarization: Recent Advances with Deep Learning." Computer Speech & Language 72, 2022. arxiv.org/...2101.09624
16. ^Injy Hamed et al. "Benchmarking Evaluation Metrics for Code-Switching Automatic Speech Recognition." arXiv:2211.16319, 2022. arxiv.org/...2211.16319