Google has released Gemini 3.5 Transcribe, a new speech-to-text model that directly tackles one of the biggest bottlenecks in voice-driven automation: messy human speech.
Unlike conventional speech recognition models that struggle with background noise and complex jargon, Gemini 3.5 Transcribe is designed to convert raw audio into polished, formatted text. According to Artificial Analysis, the new model achieves an average Word Error Rate (WER) of 4.0% for streaming and 2.6% for non-streaming use cases. It also improves the time to final transcription by 70% compared to Google’s previous model, Chirp 3.
The model is rolling out across consumer products—powering new capabilities like the Rambler feature on Android's Gboard and in the Gemini app on macOS—and is available to developers via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
Fixing the Transcription Bottleneck
If you have ever tried to automate data entry using voice notes or call recordings, you know that raw transcripts are usually a mess. People say "um," they restart their sentences, and they use industry-specific slang.
Gemini 3.5 Transcribe addresses this natively with a feature Google calls "smart transcription." It automatically removes filler words and seamlessly handles self-corrections. For example, if a speaker says, "let’s meet Tuesday—no, Wednesday," the model understands the intent and outputs the corrected text. It also auto-formats the final output and can be fed a custom vocabulary, allowing it to accurately capture specialized jargon, postal codes, order IDs, and unique spellings.
For businesses operating globally or in diverse markets, the model automatically detects and transcribes over 85 languages, navigating regional accents, diverse dialects, and live language switches.
What This Means for SMB Operations
For small and mid-sized businesses looking to automate administrative tasks, better transcription models translate directly to fewer manual corrections and more reliable data pipelines. Gemini 3.5 Transcribe is divided into two distinct APIs that align with standard business workflows:
Real-Time Streaming Voice Apps
Through the Live API (gemini-3.5-transcribe-live), the model delivers continuous, bidirectional streaming with sub-second latency. For businesses looking to deploy interactive voice agents—such as automated scheduling assistants or real-time customer support triage—this low-latency streaming makes AI interactions feel much closer to a natural phone call.
Call Logs and Meeting Processing
For asynchronous work, the Interactions API (gemini-3.5-transcribe) handles pre-recorded audio, meeting recordings, and call logs. Crucially for sales and customer service teams, it provides multi-speaker identification, accurately attributing speech with word-level timestamps for up to three speakers (identifying more than three is currently experimental). This makes generating reliable post-call analytics and automated CRM summaries far easier.
Voice-Triggered Actions Perhaps the most significant operational upgrade is the model's function calling capability. Gemini 3.5 Transcribe can delegate complex tasks to other Gemini models. Instead of simply transcribing a user request, the system can parse the intent and trigger an action—like file analysis or image generation—based on a spoken command. Google notes this is already being utilized in surfaces like Google Antigravity, where the model pairs screen context with chat history for pinpoint accuracy.
The Takeaway
Audio processing has historically been a fragile link in business automation. Models either transcribed every stutter verbatim, leaving you to write complex scripts to clean up the text, or they misunderstood key industry terms entirely.
By handling disfluency cleanup, custom vocabulary, and speaker attribution out of the box, Gemini 3.5 Transcribe removes several layers of technical friction. For operations teams, that means faster, cheaper, and more accurate ways to turn messy conversations into structured business data.