Google Launches Gemini 3.5 Transcribe With 85+ Languages and ~$0.005-Per-Minute Speech Recognition

The model automatically recognizes more than 85 languages, supports language switching within the same session and can convert conversational speech into cleaned, formatted text rather than simply reproducing every spoken word.

The more significant detail for developers is the combination of low latency, speaker recognition, custom vocabulary and aggressive API pricing.

Two models for live and recorded audio

Google is offering Gemini 3.5 Transcribe through two endpoints:

The Live version operates through the Gemini Live API and provides sub-second streaming latency, making it suitable for captions, voice interfaces, call assistants and other applications where text must appear while someone is speaking.

The standard model processes audio files of up to one hour per request. When speaker diarization or word-level timestamps are enabled, the maximum duration falls to 30 minutes. Live sessions are currently limited to 10 minutes.

Gemini 3.5 Transcribe: key specifications

Not every feature works in both modes. Speaker diarization and word-level timestamps are available for recorded audio but not for Live transcription. Google also warns that enabling word-level timestamps can reduce transcription accuracy, while speaker attribution involving three or more speakers remains experimental.

Smart transcription removes the cleanup step

The model's Smart transcription mode is designed to output what the speaker intended rather than a completely literal transcript.

It can automatically remove filler words such as “um” and “ah,” handle repeated phrases and mid-sentence corrections, insert punctuation and format numbers or other alphanumeric information.

For example, spoken input such as:

The revenue was, um, twenty six million dollars — actually twenty seven million.

Can be converted into cleaner text such as:

The revenue was $27M.

That potentially eliminates a separate LLM cleanup step in applications that previously used one model for speech recognition and another to clean the resulting transcript. Google's documentation specifically lists filler-word removal and intent-aware alphanumeric formatting as Smart transcription capabilities.

Up to 1,000 custom vocabulary terms

Developers can provide the model with a custom vocabulary containing as many as 1,000 terms.

This is particularly useful for industries where standard speech recognition frequently fails on company names, product names, technical terminology and acronyms.

A developer could, for example, supply terms such as:

Kubernetes, BigQuery, Gemini, internal product names or industry-specific terminology.

Google notes that although the maximum is 1,000 entries, customers generally see the strongest results with vocabulary lists containing up to roughly 100 terms.

85+ languages with automatic switching

Gemini 3.5 Transcribe does not require developers to specify a single language before transcription.

It performs utterance-level language detection across 85+ supported languages and can handle code-switching during the same conversation. A speaker can therefore move between supported languages without starting a new transcription session or manually changing the language configuration.

This is particularly relevant for multilingual meetings, customer-support systems and international call centers.

Speaker recognition and timestamps

For recorded audio, Gemini 3.5 Transcribe can identify different speakers and associate text with individual participants.

Speaker diarization supports up to eight speakers, although Google currently classifies attribution involving three or more speakers as experimental. The model can also generate word-level timestamps, allowing developers to map individual words back to positions in the original recording.

Those features make the model usable for automated meeting notes, podcast indexing, subtitle generation, searchable call archives and contact-center analytics without requiring a separate diarization service.

70% faster final transcription than Chirp 3

Google is positioning Gemini 3.5 Transcribe as the successor to its earlier Chirp 3 transcription technology.

According to Google, measurements from Artificial Analysis show that time to final transcription improved by approximately 70% compared with Chirp 3. The latency improvement is important because speech recognition performance is determined not only by word accuracy but by how quickly an application can confidently finalize the transcript.

For real-time voice applications, even small delays can make interfaces feel noticeably slower.

API cost: approximately $0.005 per minute

Google is also competing aggressively on price.

For the paid Gemini API tier, Gemini 3.5 Transcribe is priced at:

• $2 per 1 million audio input tokens
• $12 per 1 million output tokens

Google estimates this corresponds to approximately:

• $0.003 per minute for audio input
• $0.002 per minute for transcription output
• ~$0.005 per minute combined

At that estimated blended rate, processing 1,000 hours of audio would cost roughly $300, before accounting for other infrastructure or services used by an application. A Free Tier is also available, although Google applies usage limits to free access.

Where Gemini 3.5 Transcribe can be used

The feature set makes the model suitable for several high-volume speech workloads:

1. Call centers: transcription, speaker separation and post-call analytics.
2. Meetings: searchable transcripts with individual speakers and timestamps.
3. Voice agents: sub-second transcription feeding spoken commands into an AI system.
4. Media: subtitles, podcast transcripts and searchable video archives.
5. Professional dictation: automatic removal of filler words and corrections before text reaches an email, document or application.
6. Specialized industries: custom vocabulary can improve recognition of technical terminology, product names and acronyms.

The Live API can already be integrated with real-time communication platforms including Agora, Fishjam, LiveKit, Pipecat, Vercel and Vision Agents, according to Google's developer documentation.

Google is already deploying the technology in its own products

Gemini 3.5 Transcribe is not limited to the developer API. Google says the underlying transcription technology is being used across products including Gboard, the Gemini app, Antigravity and Chrome.

On macOS, for example, Gemini's intelligent dictation can insert cleaned text directly at the cursor while automatically handling filler words and mid-sentence corrections. The distinction is important: Google is effectively turning Gemini 3.5 Transcribe into both an API product and an input layer for its broader Gemini ecosystem.

For developers, however, the most relevant numbers are simpler: 85+ languages, up to eight speakers, 1,000 custom vocabulary terms, sub-second live transcription and an estimated API cost of roughly half a cent per minute.

Those specifications put Gemini 3.5 Transcribe directly into the market for large-scale meeting transcription, contact-center analytics and real-time voice applications.