On August 26, 2026, Google officially released Gemini 3.5 Transcribe, opening this speech-to-text model to external developers via the Gemini API. According to Google's official blog citing independent benchmarks from Artificial Analysis, the model achieves a word error rate (WER) of 2.6% in non-streaming scenarios and 4.0% in real-time streaming scenarios, with streaming latency delivering final results within 0.4 seconds after speech ends. The model is now available on Google AI Studio and the Gemini Enterprise Agent Platform, with Artificial Analysis data showing pricing of approximately $5 per thousand minutes. The release coincides with Gemini 3.5 Live and Gemini 3.5 Live Experimental on the same day, collectively forming Google's new Gemini Audio product line, led by Diego Melendo Casado, Senior Director of Engineering for Google Gemini Audio.
Two endpoints, with differences far greater than marketing materials reveal
Gemini 3.5 Transcribe is architecturally split into two independent endpoints serving distinctly different use cases, yet the capability differences between them have not been adequately emphasized at the marketing level.
The recorded audio endpoint (gemini-3.5-transcribe) targets offline batch processing, supporting up to 1 hour of audio per request, with speaker diarization (up to 8 channels) and word-level timestamps. Once speaker diarization or timestamps are enabled, the per-request limit immediately drops from 1 hour to 30 minutes. Additionally, diarization labeling for more than 3 speakers is explicitly marked as "experimental" in Google's documentation — a substantial constraint for multi-party meeting transcription scenarios.
The real-time streaming endpoint (gemini-3.5-transcribe-live) uses WebSocket for bidirectional transmission, targeting low-latency voice applications, but the per-session limit is only 10 minutes, and it does not support speaker diarization or timestamps. This means developers who want both real-time capability and speaker labeling cannot achieve this through a single endpoint in the current version. For continuous speech scenarios exceeding 10 minutes, developers need to implement their own segment-and-reconnect logic at the application layer, increasing engineering complexity.
The custom vocabulary feature allows input of up to 1,000 domain-specific terms, with Google's documentation also noting that "it is generally recommended not to exceed 100 for best results." This practical recommended limit of 100 terms is a threshold that vertical industries needing to cover large numbers of drug names, legal terminology, or financial product codes must evaluate on their own.
Smart Transcription: from "verbatim recording" to "intent understanding"
The most differentiating capability of Gemini 3.5 Transcribe lies not in language coverage, but in the processing layer Google calls "Smart Transcription." Traditional speech recognition outputs verbatim, preserving speaker self-corrections, verbal tics, and unstructured spoken language as-is; Gemini 3.5 Transcribe, by contrast, inserts a layer of semantic understanding into the transcription process.
Specifically: the model can recognize speaker self-corrections (processing "Tuesday — no, meet on Wednesday" as the final intent rather than two parallel segments); automatically filter filler words such as "um" and "uh"; and format spoken language into structured text, including correctly handling alphanumeric mixed entities such as postal codes and order numbers. This capability has been internally validated in Google's own products — the Rambler voice input feature in Gboard for Android and the Gemini app for macOS are both powered by this model. Large-scale internal deployment is a credibility signal that external developers can reference beyond laboratory data.
There is another capability currently in restricted status: Function Calling can delegate complex tasks to other Gemini models (such as image generation and file analysis). Google's documentation shows this feature is currently only available in the Gemini macOS application and has not yet been opened to API developers.
Competitive landscape: mid-tier entry, not the optimal choice
For existing speech-to-text startups such as Deepgram and AssemblyAI, this Google release applies direct pressure at the core of their business territory: low-latency real-time streaming, speaker diarization, and custom vocabulary — all three capabilities are covered. In terms of pricing, according to Artificial Analysis data, AssemblyAI's streaming endpoint is approximately $7.5 per thousand minutes, Deepgram's streaming service is approximately $7.7 per thousand minutes, and Google's $5 is competitive in the mid-tier product segment.
But the competitive landscape is not a one-sided Google rout. Artificial Analysis's non-streaming leaderboard on the release date shows Gemini 3.5 Transcribe ranked fifth with a 2.6% word error rate: ahead of it, ElevenLabs Scribe v2 has a 2.2% WER at $3.67 per thousand minutes — leading Google in both dimensions: higher accuracy and lower price. Microsoft MAI-Transcribe-1.5 ranks third with a 2.4% WER, and although its $6 pricing is higher than Google's, its accuracy is also superior.
Google's own longitudinal comparison data is more compelling: according to Artificial Analysis data, compared with Google's previous-generation Chirp 3 model, Gemini 3.5 Transcribe compresses the time from "end of speech to final transcription output" by 70%. This figure reflects the generational improvement within the same vendor, and for existing users already on Google's voice services, the upgrade value is clear.
The usage conditions behind the accuracy numbers
The 2.6% and 4.0% figures cited in Google's blog both come from Artificial Analysis's standardized benchmark evaluation, which covers approximately 8 hours of audio, weighted across three datasets. This is the most independent and transparent evaluation source currently available. However, there is a gap between the recording quality, language distribution, and domain coverage represented by the benchmark numbers and specific business scenarios — the density of medical terminology in clinical consultations, background noise in customer service recordings, and overlapping accents in multi-party meetings can all cause actual word error rates to deviate significantly from the benchmark.
This is also why Google's custom vocabulary feature is included as a base capability at the pricing level rather than a paid add-on — the quality of domain adaptation often determines real-world experience more than generic benchmark numbers in actual deployments.
Google's real strategy: the API is the entry point, Chrome is the endgame
Judging from Google's public actions, opening the API is not the core strategic goal of this release, but rather one component of a larger plan. Google has announced it will bring Gemini 3.5 Transcribe to the Chrome browser, where users will be able to input content via voice directly in any web page text field — covering everyday scenarios such as replying to messages, writing posts, and issuing commands to AI.
Chrome has long maintained a market share of over 60% in the global browser market. If voice input is seamlessly embedded at the Chrome level, the barrier to changing user input habits will be significantly lowered. For Google, the strategic logic of this path is to elevate voice from "a feature within specific applications" to "an interaction layer at the operating system level." The real-world audio data generated by every voice interaction can, in theory, feed back into improving model accuracy, forming a closed data flywheel.
The same-day release of Gemini 3.5 Transcribe and Gemini 3.5 Live also confirms Google's direction of building a complete voice infrastructure: the former focuses on transcription accuracy, the latter on real-time conversation, jointly covering different links in the "speak → understand → execute" chain.
Actionable recommendations for developers and enterprises
The following are practical recommendations based on currently known information:
- Rapid prototyping and validation scenarios: Google AI Studio offers free credits, suitable for testing accuracy with your own actual audio samples rather than making selection decisions based solely on benchmark numbers. Comparative experiments on domain-specific audio (medical, legal, customer service, etc.) are especially recommended.
- Real-time voice applications: The 10-minute session limit is a hard constraint of the current version. For scenarios exceeding this limit, evaluate the engineering cost of segment-and-reconnect before deciding whether to adopt the real-time streaming endpoint.
- Meeting transcription/call analysis: The batch processing endpoint performs most stably in scenarios with ≤2 speakers and no need for real-time output; for more than 3 speakers, confirm whether experimental labeling meets business tolerance.
- Enterprises with existing vendor contracts: The Artificial Analysis leaderboard shows ElevenLabs Scribe v2 outperforms Google's current version in both accuracy and price. It is worth evaluating side-by-side before switching vendors, rather than defaulting to a platform premium.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接