Alphabet Inc. subsidiary Google released Gemini 3.5 Transcribe, a dedicated audio model designed to convert spoken thoughts directly into structured, formatted text. Replacing the earlier Chirp 3 speech engine, the new system filters out natural speech pauses and filler words such as “um” and “ah” while applying correct punctuation and sentence structure on the go. It operates across more than 85 languages and handles complex regional accents, alphanumeric strings like postal codes, and mid-sentence corrections.
Key Takeaways
- Converts unstructured conversational speech into clean, formatted text without filler words.
- Achieves a 4.0 percent Word Error Rate in streaming mode and 2.6 percent in non-streaming tests.
- Supports over 85 languages with automated speaker attribution for up to three people in recorded audio.
- Integrates directly into Gboard on Android, Google Antigravity, Google AI Studio, and Gemini for macOS.
- Connects with other Gemini models through function calling to execute desktop tasks via voice.
High Accuracy and Low Latency Across Devices
According to independent benchmark data from Artificial Analysis, Gemini 3.5 Transcribe records a 70 percent faster time to final transcription compared to previous generation systems. The model logs an average Word Error Rate (WER) of 4.0 percent for live streaming audio and 2.6 percent for pre-recorded audio files. On the multilingual FLEURS benchmark, it scored a 5.50 percent streaming WER and a 5.04 percent non-streaming WER.
The model specifically tackles real-world speech hurdles. It accurately parses technical jargon, custom vocabulary, and alphanumeric entities like flight numbers and tracking codes that typically confuse standard speech recognition software. When processing pre-recorded files through Google AI Studio or the Gemini API, it provides word-level timestamps and separates distinct speakers.
Everyday Integrations from Gboard to Web Browsers
Google is embedding Gemini 3.5 Transcribe across its consumer and developer software. On Android devices, the model runs inside Gboard through a new feature called Rambler. This lets users dictate long thoughts naturally, while the software formats paragraphs, corrects slips, and adjusts tone through follow-up voice commands.
On macOS, the standalone Gemini desktop app pairs the model with screen context. Users can dictate instructions to summarize files or generate pictures at their active cursor position. Developers can access the model in Google AI Studio and Google Antigravity to build voice agents and dictate code in real time. Google also confirmed a coming update for the Chrome web browser that will allow direct talk-to-type input inside any online text field.
Frequently Asked Questions
Q1. What is Gemini 3.5 Transcribe?
A1. It is a specialized audio AI model from Google built to transcribe real-time speech and recorded audio into structured text while removing verbal disfluencies like pauses and filler words.
Q2. How many languages does Gemini 3.5 Transcribe support?
A2. The model automatically detects and transcribes speech across more than 85 languages and regional dialects.
Q3. How accurate is Gemini 3.5 Transcribe?
A3. Artificial Analysis testing indicates an average Word Error Rate of 4.0 percent in live streaming mode and 2.6 percent for non-streaming audio tasks.
Q4. Where can users access Gemini 3.5 Transcribe?
A4. The model is available via Gboard on Android, Gemini for macOS, Google AI Studio, the Gemini API, and will roll out to Google Chrome text fields.
