Google: Gemini 3.5 Transcribe

Google Gemini 3.5 Transcribe is the synchronous model for converting complete, prerecorded audio files into written speech. It is based on Gemini’s audio-understanding capabilities, but has a transcription-specific input and output contract. The separate Gemini 3.5 Transcribe Live edition receives an audio stream and returns incremental results. This recorded-audio model produces a completed transcript, with optional word timing or speaker attribution depending on the selected configuration.

Prepare a clear recording with speech that is not severely clipped and with meaningful separation between voices where possible. The manufacturer’s synchronous examples pass WAV audio as file bytes or a stored audio resource, with its MIME type specified. Keep the original audio available for review. The intended result might be a checked interview transcript, a caption-alignment record or readable dictation; decide which artifact is needed before combining features that change the wording or add metadata.

Google documents language auto-detection across more than 85 languages, including code-mixing within a session. Language support levels vary, and a known language hint can improve recognition. Custom vocabulary supplies distinct names, brands or specialist terms that are likely to occur. The maker allows up to 1,000 terms and notes that focused lists of up to about 100 typically work best. Keep uncommon terms relevant to the recording and check them against the audio rather than assuming a vocabulary hint proves the term was spoken.

VERBATIM is the default mode. It preserves fillers, repetitions, false starts and spoken corrections so the text remains a record of the wording. This is the mode to use when the task needs word-level timestamps or speaker diarization. For an interview, for example, retain the original turns and timings, then correct important names and quantities by listening to the relevant passage. A later summary or polished article can use that checked record without silently replacing what the speaker said.

SMART mode is designed for readable dictation. It removes disfluencies, resolves inline self-corrections and structures spoken material into paragraphs or lists, with normalized dates and numbers where appropriate. A dictated task list can become a readable list rather than a string of spoken transitions. That processing changes the form of the transcript. Google explicitly states that SMART is incompatible with word timestamps and diarization, so choose it when readable text is the required artifact rather than attempting to use it as the timed, speaker-attributed record.

When word timing is enabled, the response includes recognized words with start and end offsets. These support locating a correction in the recording or preparing captions. The maker identifies an accuracy trade-off from enabling word-level timestamps. Review the text as well as the alignment, particularly around fast speech, overlap or unusual terminology. A useful timed transcript retains the word record and the source audio together instead of treating a fluent sentence as proof that its alignment is correct.

Speaker diarization identifies and labels voices in recorded audio, with up to eight speakers documented. Attribution for three or more speakers is described by the maker as experimental. Treat the labels as inferred speaker groupings and review transitions against the recording. A speaker label is not independent proof of a person’s identity or role. If the task needs names, associate the labels with known recording information rather than inventing identities from the recognized words.

The manufacturer’s feature table lists a 15-minute recorded-audio window, with file processing limited to that duration when features such as timestamps or diarization are enabled. Prepare longer source material as suitable segments while preserving their order and original time position. The model’s transcript, timed words and speaker-labeled parts are different result representations; retain the representation needed for the task. These audio-specific limits should not be replaced with context or duration figures from a general Gemini multimodal model. Google lists SMART formatting, word-level timestamps and speaker diarization as experimental features.

The output contract is transcription-specific. SMART formatting creates readable text but does not add a general JSON-schema response contract, tool-calling capability or an instruction-following assistant workflow. For the final artifact, check important wording, numbers and negations against the audio, preserve any unresolved segment for correction, and retain the selected mode with the result. This makes it clear whether the deliverable records the original spoken wording or presents an edited reading version derived from it.