Gemini 3.5 Transcribe Release Notes

Gemini 3.5 Transcribe provides intelligent, real-time speech-to-text conversion

Google has released Gemini 3.5 Transcribe, a speech-to-text (STT) model designed to convert raw audio into polished, formatted text. The model distinguishes itself from conventional speech recognition by automatically removing filler words ("ums" and "ahs"), handling self-corrections (e.g., "let’s meet Tuesday—no, Wednesday"), and auto-formatting output to capture user intent rather than just literal speech.

API Availability and Implementation

Gemini 3.5 Transcribe is available to developers through two distinct APIs to support different architectural needs:

Real-time streaming (gemini-3.5-transcribe-live): Accessible via the Live API, this version provides bidirectional streaming with sub-second latency for interactive voice applications.
Pre-recorded audio processing (gemini-3.5-transcribe): Accessible via the Interactions API, this version handles recorded audio, meeting logs, and call records, providing speaker attribution for up to three speakers and word-level timestamps.

Developers can access these models via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

Technical Performance and Benchmarks

Gemini 3.5 Transcribe demonstrates significant improvements over its predecessor, Chirp 3, specifically in latency and accuracy:

Word Error Rate (WER): According to Artificial Analysis, the model achieves an average WER of 4.0% for streaming and 2.6% for non-streaming use cases. On the FLEURS benchmark, it achieves a 5.50% WER in streaming and 5.04% in non-streaming modes.
Latency: Time to final transcription has improved by 70% compared to Chirp 3.
Language Support: The model automatically detects and transcribes over 85 languages, including regional accents and diverse dialects.
Customization: It supports custom vocabulary to recognize specialized jargon and unique spellings.

Integration Across Google Ecosystem

Google is integrating the model into several consumer and developer surfaces:

Gboard (Android): The "Rambler" feature uses 3.5 Transcribe to transform spoken thoughts into formatted text and allows voice-based edits and style changes.
Gemini App (macOS): Enables natural speech transcription and voice commands that pair with screen context to summarize local files or generate images.
Google Antigravity: Pairs screen context and chat history to improve transcription accuracy for file names and active documents.
Chrome: A forthcoming feature will allow users to "talk to type" in any web field.

Community Feedback and Technical Critiques

While the official release highlights performance gains, developer and user discussions on Hacker News reveal several points of contention and technical gaps:

Accuracy vs. "Smart" Editing

Some users report that the "smart transcription" feature, which removes disfluencies, can inadvertently remove meaningful context. One user noted that phrases like "I hesitated to check it" were stripped out, altering the intended meaning of the sentence.

Diarization and Latency

Critics point out that speaker diarization is limited to three speakers (and is experimental for more), which lags behind competitors like Soniox and Deepgram. Additionally, some developers argue that while accuracy is high, latency still requires improvement for high-performance real-time translator apps.

Reliability and Hallucinations

Users expressed concern over whether the model inherits hallucination issues from Chirp. One developer mentioned that Chirp would sometimes generate repetitive text (e.g., "I don't know. I don't know.") when encountering noise or silence, leading them to use a hybrid approach of Whisper for timestamps and Gemini Flash for correction.

Pricing and Accessibility

Community members highlighted the lack of transparent pricing in the announcement. Based on API documentation, the costs are as follows:

Model	Input Cost	Output Cost
3.5 Transcribe Live	$3.50 / 1M tokens (or $0.005/min audio)	$21.00 / 1M tokens (or $0.004/min text)
3.5 Transcribe	$2.00 / 1M tokens (or $0.003/min audio)	$12.00 / 1M tokens (or $0.002/min text)

Tooling and Workflow

Some developers reported friction with Google Cloud billing account processing times, noting that setup can take significantly longer than competitors like OpenRouter.