Google has introduced Gemini 3.5 Transcribe, a new speech-to-text model designed to make voice interactions with Gemini more accurate, responsive, and useful. Unlike basic transcription systems that simply turn spoken words into text, the model is designed to understand how people actually speak, including corrections, filler words, accents, and specialized terminology.
That distinction matters because better speech recognition can directly improve how an AI assistant understands a user’s request. If Gemini receives a cleaner and more accurate version of what someone said, it has a better chance of responding to the actual intent rather than trying to interpret transcription mistakes.
Gemini 3.5 Transcribe is built for natural speech
People rarely speak in perfectly structured sentences. Someone might say, “Let’s meet Tuesday — no, Wednesday,” or pause repeatedly while thinking about what to say.
Google says Gemini 3.5 Transcribe can handle these situations by removing filler words such as “um” and “ah,” recognizing self-corrections and automatically formatting the resulting text. The goal is to turn messy, conversational speech into text that more closely represents what the person intended to say.
The model can also recognize custom vocabulary. That could help when someone talks about technical terms, product names, file names, or other words that ordinary speech-recognition systems may misinterpret.
Google says the model supports more than 85 languages and can automatically detect languages while handling regional accents and dialects. It can also recognize multiple speakers in recorded audio, with speaker attribution and timestamps for up to three speakers.
The accuracy improvements could matter more than the name
Google reports a Word Error Rate of 4.0% for streaming transcription and 2.6% for non-streaming use cases, based on measurements from Artificial Analysis. In simple terms, Word Error Rate measures how often the system mishears spoken words, so a lower number generally means more accurate transcription.
The model also improves on Google’s previous Chirp 3 transcription system. Google says time to final transcription has improved by 70%, which matters for voice conversations where waiting for text to appear can make an assistant feel slow.
For Gemini, lower latency is just as important as accuracy. A transcription system that understands every word but takes too long to process them would still make a voice conversation feel unnatural.
Gemini 3.5 Transcribe is designed for real-time streaming with sub-second latency through Google’s Live API. Recorded audio can also be processed through the Interactions API, with features such as speaker identification and word-level timestamps.
Gemini could do more with what users say
The biggest change may happen after speech is transcribed.
Google says Gemini 3.5 Transcribe can work with other Gemini models through function calling. That means spoken input does not have to stop at transcription. Gemini can use the recognized request to trigger other actions, such as analyzing a file or generating an image.
Google is already using the technology in several products. On Android, the Rambler feature in Gboard can turn spoken thoughts into formatted text, remove filler words, and let users make corrections or change the writing style using their voice.
On the Gemini app for macOS, Google says users can use natural speech to interact with files and other on-screen content. A spoken request could, for example, ask Gemini to summarize a local file, reuse text in another application, or generate an image at the cursor.
That points toward a broader change in how Gemini could handle dialogue. Instead of treating voice as simply another way to enter a text prompt, Gemini can use speech, screen context and conversation history together to determine what the user is trying to accomplish.
Voice recognition is becoming part of the Gemini experience
Google is also bringing Gemini 3.5 Transcribe to developer platforms, allowing companies to build voice agents, live captioning systems and tools for analyzing recorded calls. The real-time version is available through the Live API, while recorded audio uses the Interactions API.
For ordinary users, the technology could eventually become most visible in everyday tasks. Google says talk-to-type functionality is coming to Chrome, where users can dictate replies, draft posts, and enter prompts into web fields using their voice.
The important shift is that Gemini may no longer need users to speak in a way that resembles written language. If the system can clean up corrections, understand specialized words, and retain the meaning of conversational speech, users can simply talk.
That could make Gemini’s voice capabilities feel less like dictation and more like an actual dialogue.
The remaining question is how widely Google will deploy these capabilities. Gemini 3.5 Transcribe is already in public preview for developers and enterprises, while some consumer features are available on Android and macOS, with Chrome support still listed as coming soon.
As voice becomes a larger part of Gemini, the quality of the first few seconds of recognition may matter just as much as the model generating the final answer. Gemini 3.5 Transcribe is Google’s attempt to make that first step considerably more reliable.
