Google has introduced Gemini 3.5 Transcribe, a new speech-to-text (STT) model designed to deliver more accurate, intelligent and natural transcription across real-time and recorded-audio applications.
Announced on August 26, 2026, Gemini 3.5 Transcribe is designed to convert raw speech into accurate, polished and formatted text while handling challenges such as background noise, filler words, self-corrections, specialized terminology and different accents.
According to Google, the model achieves an average Word Error Rate (WER) of 2.6% for non-streaming transcription and 4.0% for streaming use cases, based on measurements from Artificial Analysis. It also supports automatic detection and transcription across more than 85 languages.
Gemini 3.5 Transcribe Offers Two Different Experiences
One of the key features of Gemini 3.5 Transcribe is that Google has designed separate capabilities for real-time and pre-recorded audio.
Real-Time Streaming: Built for Voice Agents
The streaming version is designed for applications where response speed is critical. It provides continuous, bidirectional audio streaming with sub-second latency through Google’s Live API.
This makes it suitable for applications such as:
- AI voice assistants
- Customer-service voice agents
- Real-time conversational applications
- Live captioning
- Voice-controlled systems
- Interactive call-center applications
Google reports an average 4.0% WER for streaming transcription. The focus here is on responsiveness, allowing applications to process speech while a conversation is taking place.
For developers building real-time voice agents, this could be particularly important because even small improvements in transcription accuracy and latency can make conversations feel considerably more natural.
Non-Streaming Transcription: Accuracy and Rich Features
For recorded audio, meetings, interviews, podcasts, and call recordings, Gemini 3.5 Transcribe offers a more feature-rich transcription experience.
Google reports an average 2.6% WER for non-streaming transcription, making this the stronger option when accuracy and detailed transcript processing are more important than real-time response speed.
The model can also provide speaker attribution and word-level timestamps for recorded audio. Google says speaker identification can support up to three speakers, while support for more than three speakers remains experimental.
This makes the non-streaming version useful for:
- Meeting transcription
- Podcast and interview transcription
- Video subtitles
- Call-center analysis
- Content creation
- Searchable audio archives
- Business intelligence
- Recorded customer conversations
What Makes Gemini 3.5 Transcribe Different?
Gemini 3.5 Transcribe goes beyond simply converting spoken words into text.
Google describes several intelligent transcription capabilities that can make the final transcript more useful without requiring extensive manual editing.
Smart Transcription
The model can handle common characteristics of natural speech, including filler words, self-corrections, and disfluencies.
For example, if someone says, “Let’s meet Tuesday—no, Wednesday,” the model can understand the correction and produce a cleaner transcript reflecting the intended statement.
It can also remove filler words such as “um” and “ah” and automatically format the resulting text.
This is particularly useful for content creators, businesses, and developers who need a polished transcript rather than a raw word-for-word audio dump.
Custom Vocabulary for Specialized Applications
Another important feature is custom vocabulary support.
Businesses often use terminology that general-purpose speech recognition systems may struggle to recognize, including product names, technical terms, company names, medical terminology, financial terminology and order IDs.
Gemini 3.5 Transcribe allows developers to provide custom vocabulary so the model can better recognize specialized terms and unusual spellings.
This could be especially valuable for enterprise applications where transcription accuracy needs to extend beyond everyday conversations.
Support for More Than 85 Languages
Google is also positioning Gemini 3.5 Transcribe as a multilingual speech model.
The system can automatically detect and transcribe speech across more than 85 languages, with support for regional accents and dialects. Google’s documentation lists a broad range of supported language and locale combinations, including Indian languages such as Bengali, Kannada, and Assamese.
For global businesses, this means voice applications can potentially be deployed across multiple markets without requiring a completely separate speech-recognition system for every language.
70% Faster Final Transcription Than Chirp 3
Google also highlights a significant performance improvement over its previous-generation Chirp 3 technology.
According to Google, Gemini 3.5 Transcribe delivers a 70% improvement in time-to-final-transcription compared with Chirp 3.
This could be important for applications that process large volumes of recorded audio, where the time required to produce the final transcript directly affects workflow efficiency.
Real-World Applications
The combination of improved accuracy, multilingual support, and intelligent transcript processing opens up several potential applications.
Customer Service and Voice Agents
AI-powered customer-service systems can use the streaming model to understand customers in real time. Better transcription can improve intent detection, reduce misunderstandings and enable more natural conversations.
Content Creation
Podcasters, YouTubers and video producers can use automated transcription to generate subtitles, searchable transcripts and content drafts from recorded audio.
Meetings and Productivity
Companies can automatically transcribe meetings and conversations, identify speakers and generate structured information from recorded discussions.
Accessibility
Accurate speech recognition can improve live captions and make spoken content more accessible to people who rely on text-based communication.
Healthcare and Legal Workflows
Recorded consultations, interviews, and other professional conversations can potentially be converted into searchable transcripts, although organizations in regulated industries will still need to consider privacy, compliance, and human-review requirements.
Business Analytics
Call recordings and customer conversations can be transformed into text that can then be analyzed for recurring issues, customer feedback, product demand and operational trends.
Gemini 3.5 Transcribe and the Developer Opportunity
For developers, perhaps the most important aspect of the release is the availability of the model through APIs.
Google has separated the real-time and recorded-audio use cases, allowing developers to select the architecture that matches their application.
A voice-agent developer may prioritize the streaming version and low latency, while a company building a meeting-transcription or call-analysis platform may prioritize the non-streaming model and its richer transcription features.
This gives developers greater flexibility instead of forcing the same speech-recognition configuration onto every application.
Accuracy Numbers Need to Be Viewed in Context
The headline 2.6% WER is impressive, but it is important to understand what the number represents.
Google’s announcement states that the 2.6% non-streaming and 4.0% streaming figures were measured by Artificial Analysis. Google also reports separate results on the FLEURS multilingual benchmark, where it records 5.04% WER for non-streaming and 5.50% for streaming.
Therefore, the 2.6% figure should not automatically be interpreted as a universal error rate for every language, accent, microphone or real-world recording.
Actual transcription quality can vary depending on audio quality, background noise, speakers, accents, terminology and the specific application.
Google’s Broader AI Strategy
Gemini 3.5 Transcribe is another indication that Google is expanding the Gemini ecosystem beyond traditional text-based AI.
The model is already being used in Google products, including new voice capabilities in the Gemini app and Android, according to Google. The company is also positioning the technology for developers building voice-enabled applications.
The ability to combine speech recognition with other AI capabilities could eventually make voice interfaces significantly more powerful, allowing applications not only to understand what users say but also to take actions based on those conversations.
Gemini 3.5 Transcribe: Key Takeaways
Google’s latest speech model brings together several important capabilities:
- 2.6% average WER for non-streaming transcription
- 4.0% average WER for streaming transcription
- 85+ languages with automatic language detection
- Real-time streaming for interactive voice applications
- Speaker attribution and word-level timestamps for recorded audio
- Custom vocabulary support
- Automatic removal of filler words
- Handling of self-corrections and disfluencies
- Automatic formatting of transcripts
- 70% faster time-to-final-transcription compared with Chirp 3, according to Google
Points to consider
Gemini 3.5 Transcribe represents a significant step forward in Google’s speech AI strategy.
Rather than treating speech-to-text as a simple conversion from audio to words, Google is building a system that attempts to understand natural speech and produce cleaner, more useful text.
The 2.6% non-streaming WER, 4.0% streaming WER, support for 85+ languages, custom vocabulary, and intelligent transcript processing make Gemini 3.5 Transcribe an important development for voice agents, content creation, business communication, and AI-powered applications.
For developers, the biggest opportunity may be the combination of transcription with the broader Gemini ecosystem. Voice interfaces can increasingly move from simply “hearing” users to understanding their requests and taking actions.
As speech becomes a more important interface for AI applications, Gemini 3.5 Transcribe puts Google in a stronger position in the rapidly developing voice-AI and speech-recognition market.