Gemini 3.5 Transcribe: Google’s New Speech-to-Text Model for Audio Files
Google has introduced Gemini 3.5 Transcribe, a dedicated speech-to-text model built on Gemini’s audio understanding capabilities with language detection, speaker diarization, word-level timestamps, and custom vocabulary.
Google has introduced Gemini 3.5 Transcribe, a dedicated speech-to-text model built on Gemini’s audio understanding capabilities. The model is designed for accurate, low-latency transcription from audio files and includes automatic language detection, speaker diarization, word-level timestamps, custom vocabulary biasing, and smart transcription.
The official model ID is:
gemini-3.5-transcribe
Google announced Gemini 3.5 Transcribe as generally available on August 26, 2026. A separate model, gemini-3.5-transcribe-live, is intended for low-latency streaming transcription over WebSockets. This article focuses on the non-streaming audio-file model.
What makes Gemini 3.5 Transcribe different?
Gemini 3.5 Transcribe is more than a basic speech-to-text endpoint. It is designed to handle real-world audio with multiple languages, speaker changes, background noise, and domain-specific terminology.
The model automatically detects speech across 85+ languages and can handle code-switching within and between sentences. It can also identify speakers, return word-level timing, and apply smart formatting to make the output easier to read.
| Capability | What it provides |
|---|---|
| Automatic language detection | Detects supported languages without manual setup |
| Code-switching | Handles conversations that move between languages |
| Speaker diarization | Separates speakers in audio-file transcription |
| Word-level timestamps | Returns start and end offsets for recognized words |
| Custom vocabulary | Biases recognition toward names, jargon, brands, and acronyms |
| Smart transcription | Removes disfluencies and formats numbers, punctuation, and layout |
| Normalized output | Formats numbers, dates, currency, and spoken conventions |
File limits and feature tradeoffs
Gemini 3.5 Transcribe accepts audio files up to one hour per request under the general model limit. When speaker diarization or word-level timestamps are enabled, the documented file-processing limit is 30 minutes.
Speaker diarization supports up to eight speakers. Attribution for recordings with three or more speakers is described as experimental, so meeting transcripts should still be reviewed before publication or archival use.
| Requirement | Recommended approach |
|---|---|
| Standard file transcription up to 1 hour | Send as a single request |
| Transcription requiring speaker diarization | Split audio into segments of 30 minutes or less |
| Transcription requiring word-level timestamps | Split audio into segments of 30 minutes or less |
| Long meetings or multi-hour files | Segment by time, request transcription, and merge output |
| Live microphone or streaming audio | Use gemini-3.5-transcribe-live instead |
When splitting long recordings, include a slight overlap at segment boundaries to avoid cutting off words, and verify speaker labels after merging.
Create a Gemini API key
Open Google AI Studio, sign in, and create or view a Gemini API key. Google’s current documentation recommends using environment variables such as GEMINI_API_KEY or GOOGLE_API_KEY rather than hard-coding keys in source code.
export GEMINI_API_KEY='YOUR_GEMINI_API_KEY'
Google’s API-key documentation states that newly created AI Studio keys are authorization keys and that standard-key handling is changing during 2026. Apply appropriate restrictions and follow the migration guidance shown in your Google AI Studio or Google Cloud account.
Treat the key like a password. Never publish it in a blog post, Git repository, screen recording, or application distributed to end users.
Transcribe an audio file with Python
Install the Google GenAI SDK:
pip install google-genai
The official transcription guide recommends uploading an audio file and passing its URI to the Interactions API. A minimal Python example is:
from google import genai
client = genai.Client()
audio_file = client.files.upload(file="sample.mp3")
interaction = client.interactions.create(
model="gemini-3.5-transcribe",
input=[
{
"type": "audio",
"uri": audio_file.uri,
"mime_type": audio_file.mime_type,
}
],
)
print(interaction.output_text)
With GEMINI_API_KEY set in the environment, the client library can authenticate without placing the key directly in the script.
Automatic language detection and language hints
Automatic language detection is the default. You can omit language codes when you want Gemini to identify the spoken language automatically and handle multilingual switching.
If the language is known in advance, provide BCP-47 language codes through the transcription configuration. For an English recording, examples include:
generation_config = {
"transcription_config": {
"language_codes": ["en-US"],
}
}
Language hints are useful when the recording contains a dominant language, regional vocabulary, or names that could otherwise be interpreted in multiple ways. For multilingual recordings, compare automatic detection with explicit hints rather than assuming one setting will be best for every file.
Custom vocabulary for names and terminology
Custom vocabulary biasing helps the model recognize uncommon words, technical jargon, brand names, acronyms, and proper nouns. The API accepts up to 1,000 phrases, although Google’s documentation says that the best results are typically achieved with up to 100 terms.
interaction = client.interactions.create(
model="gemini-3.5-transcribe",
input=[
{
"type": "audio",
"uri": audio_file.uri,
"mime_type": audio_file.mime_type,
}
],
generation_config={
"transcription_config": {
"custom_vocabulary": [
"Kubernetes",
"BigQuery",
"Scribis",
"Speech-to-Text",
],
}
},
)
Vocabulary is a recognition hint, not a command to insert terms that were never spoken. Compare results with and without the vocabulary list, and remove terms that are unrelated to the recording.
Smart transcription and formatting
Smart transcription is designed to make raw speech easier to read. It can remove filler words, repetitions, and some disfluencies, while applying capitalization, punctuation, structured formatting, and inverse text normalization.
This is useful for meeting notes, interviews, and articles, but a polished transcript is not always a verbatim transcript. If you need a legal, research, or archival record, compare the smart output with a less-processed version and decide which details must be preserved.
Speaker diarization and timestamps
Audio-file transcription supports both speaker diarization and word-level timestamps. These features can turn a recording into a more useful foundation for interviews, panels, meetings, and captions.
They also reduce the maximum file duration to 30 minutes, so long recordings may need to be split before processing. When splitting a recording, keep a small amount of context around segment boundaries and check for duplicated or missing words when merging the results.
Pricing and free-tier data use
Google’s current Gemini Developer API pricing page lists the following reference values for gemini-3.5-transcribe:
| Tier | Audio input | Text output | Data-use label |
|---|---|---|---|
| Free | Free of charge | Free of charge | Content used to improve Google products |
| Paid Standard | US$2.00 per 1M audio tokens, estimated US$0.003/minute | US$12.00 per 1M text tokens, estimated US$0.002/minute | Content not used to improve Google products |
Google estimates an effective blended rate of approximately US$0.005 per minute, based on 25 audio tokens per second and 175 text tokens per minute.
The free tier is not simply a cheaper version of the paid tier from a data-policy perspective. The current pricing page explicitly labels free-tier content as used to improve Google products and paid-tier content as not used to improve Google products. Review the current terms and project settings before uploading confidential interviews, internal meetings, customer calls, or other sensitive recordings.
Artificial Analysis Benchmark Data
Beyond official specifications, independent AI evaluation platform Artificial Analysis has benchmarked Gemini 3.5 Transcribe in its Speech-to-Text evaluations.
The evaluation measures accuracy via the Artificial Analysis Word Error Rate index (AA-WER v2, covering real-world speech with diverse accents and acoustic conditions), processing speed factor, and pricing per 1,000 audio minutes:
| Model & Provider | AA-WER (Lower is better) | Speed Factor (Higher is faster) | Price / 1,000 mins (USD) |
|---|---|---|---|
| Gemini 3.5 Transcribe (Google) | 2.6% | 79.6x | $5.00 |
| Scribe v2 (ElevenLabs) | 2.2% | 53.9x | $3.67 |
| MAI-Transcribe-1.5 (Microsoft Azure) | 2.4% | 191.5x | $6.00 |
| GPT Transcribe (OpenAI) | 3.3% | 40.8x | $4.50 |
| Nova-3 (Deepgram) | 5.2% | 540.2x | $4.30 |
| Whisper Large v3 Turbo (Groq) | 4.6% | 165.4x | $0.67 |
According to benchmark results, Gemini 3.5 Transcribe delivers top-tier transcription accuracy with an AA-WER of 2.6%, outperforming models such as OpenAI's GPT Transcribe (3.3%) and Deepgram's Nova-3 (5.2%). Operating at a 79.6x real-time speed factor, it achieves an impressive balance of high precision, fast processing, and competitive pricing.
Add Scribis to the review stage
Cloud ASR is excellent at processing audio, but transcript work often continues afterward. You may need to correct names, search for a quote, remove a repeated phrase, split an interview into sections, or prepare subtitles.
Scribis can be used as a desktop workspace for that editorial stage. Bring the recording and transcript into a focused environment, review the text, organize sections, and export the result as notes or subtitles.
This creates a practical workflow: Gemini 3.5 Transcribe handles speech recognition, while Scribis helps turn the result into a transcript that is easier to review and use.
Gemini 3.5 Transcribe versus Gemini 3.5 Transcribe Live
These models serve different purposes:
| Model | Input style | Best for |
|---|---|---|
gemini-3.5-transcribe |
Uploaded audio files | Recorded meetings, interviews, lectures, podcasts, and captions |
gemini-3.5-transcribe-live |
Bidirectional audio streaming over WebSockets | Microphone input, live captions, and real-time applications |
Do not use the Live model’s streaming workflow or pricing as a substitute for file transcription. Google documents them as separate model endpoints with different limits and request patterns.
Common questions
Can it transcribe English accents?
Gemini 3.5 Transcribe is designed for diverse accents and background conditions, but real-world accuracy still depends on microphone quality, noise, vocabulary, speaker overlap, and the recording environment. Test it with recordings that reflect your actual audience.
Can I process a one-hour meeting with speaker labels?
The general model limit is one hour, but the documented file limit becomes 30 minutes when speaker diarization or word-level timestamps are enabled. Split the meeting into shorter segments if those features are required.
Is the free tier private?
The current Gemini API pricing page labels free-tier content as used to improve Google products and paid-tier content as not used to improve Google products. Free access should therefore not be treated as equivalent to a paid privacy configuration or zero-retention service.
Can I use a standard API key forever?
Google’s API-key documentation says that key handling is changing during 2026 and that standard keys are scheduled for rejection in September 2026. Use the current AI Studio migration guidance and create an authorization key when available.
Conclusion
Gemini 3.5 Transcribe is a capable new Cloud ASR model for English and multilingual audio. Its strongest features are automatic language detection, code-switching, speaker diarization, word-level timestamps, custom vocabulary, and smart formatting.
Start with a plain English audio-file request, then add language hints, vocabulary, timestamps, or speaker labels as your workflow requires. After the API produces the transcript, Scribis can provide a practical desktop environment for reviewing, editing, searching, organizing, and exporting spoken content.