{"id":2172,"date":"2026-08-27T05:30:32","date_gmt":"2026-08-27T12:30:32","guid":{"rendered":"https:\/\/www.five.reviews\/?p=2172"},"modified":"2026-08-27T05:30:35","modified_gmt":"2026-08-27T12:30:35","slug":"gemini-3-5-transcribe","status":"publish","type":"post","link":"https:\/\/www.five.reviews\/ai-tools\/gemini-3-5-transcribe\/","title":{"rendered":"Gemini 3.5 Transcribe: Benchmarks, Pricing & Features"},"content":{"rendered":"\n

Gemini 3.5 Transcribe<\/a> is Google’s latest speech-to-text model, engineered to deliver intelligent, low-latency transcription with speaker attribution, word-level timestamps, and smart formatting. Unlike conventional transcription services that capture raw audio word-for-word, Gemini 3.5 Transcribe processes speech with context awareness, handling self-corrections, filler words, punctuation, and domain-specific terminology.<\/p>\n\n\n\n

Available in both recorded-audio and real-time streaming variants, the model is now in public preview through the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform. For developers building voice applications, meeting analytics systems, or multilingual transcription pipelines, this model represents a significant step forward in transcription accuracy, latency, and usability.<\/p>\n\n\n\n

Gemini 3.5 Transcribe at a Glance<\/strong><\/h2>\n\n\n\n
Aspect<\/strong><\/td>Details<\/strong><\/td><\/tr>
What it is<\/td>Google’s speech-to-text model with smart formatting and language detection<\/td><\/tr>
Model IDs<\/td>gemini-3.5-transcribe (recorded), gemini-3.5-transcribe-live (streaming)<\/td><\/tr>
Best use case<\/td>Meeting transcription, voice applications, multilingual audio processing<\/td><\/tr>
Languages supported<\/td>85+ languages with automatic detection<\/td><\/tr>
Streaming availability<\/td>Yes, via Live API with sub-second latency<\/td><\/tr>
Benchmark WER (Artificial Analysis)<\/td>4.0% streaming, 2.6% non-streaming<\/td><\/tr>
Starting price<\/td>~$0.005\/min (recorded), ~$0.009\/min (live)<\/td><\/tr>
Main differentiator<\/td>Smart transcription with intent recognition and automated cleanup<\/td><\/tr>
Main limitation<\/td>Public preview status; no speaker diarization in live mode<\/td><\/tr>
Current availability<\/td>Public preview, paid tier available<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Understanding Gemini 3.5 Transcribe vs. Transcribe Live<\/strong><\/h2>\n\n\n\n

Gemini 3.5 Transcribe comes in two distinct versions serving different use cases. This distinction is critical because the endpoints have different capabilities, limitations, and pricing.<\/p>\n\n\n\n

Gemini 3.5 Transcribe<\/strong> (gemini-3.5 Transcribe) is designed for pre-recorded audio. It accepts audio files up to 1 hour in length and processes them through the Interactions API. The recorded-audio endpoint supports speaker diarization (identifying who said what), word-level timestamps (useful for video editing and podcasting), custom vocabulary biasing, and smart transcription.<\/p>\n\n\n\n

Gemini 3.5 Transcribe Live<\/strong> (gemini-3.5-transcribe-live) handles real-time streaming audio over WebSockets. Sessions are limited to 10 minutes, but the model delivers sub-second latency for interactive applications. Live transcription does not currently support speaker diarization or word-level timestamps, but it does support custom vocabulary and smart formatting.<\/p>\n\n\n\n

The choice between these endpoints depends entirely on your workflow. If you’re processing recorded meetings, calls, or podcasts, use Transcribe. If you’re building voice agents, live caption systems, or interactive voice applications, use Transcribe Live.<\/p>\n\n\n\n

Key Takeaway:<\/strong> Gemini 3.5 Transcribe combines speech recognition with smart formatting, language detection, custom vocabulary, speaker diarization (recorded only), and word-level timestamps (recorded only). Choose recorded for files and meetings; choose live for real-time, bidirectional interactions.<\/p>\n\n\n\n

What Gemini 3.5 Transcribe Does Differently<\/strong><\/h2>\n\n\n\n
\"What<\/figure>\n\n\n\n

Smart Transcription<\/strong><\/h3>\n\n\n\n

Gemini 3.5 Transcribe is designed to produce cleaner, more useful output than a simple word-for-word transcript. The model interprets speaker intent and applies formatting automatically. When someone says, “Let’s meet Tuesday, no, Wednesday,” the smart transcription engine understands the self-correction and outputs “Let’s meet Wednesday” rather than reproducing the spoken disfluency verbatim.<\/p>\n\n\n\n

This approach extends to filler words. Conversational “um,” “uh,” “you know,” and “like” are automatically removed, producing professional-quality transcripts without manual cleanup. The model also improves punctuation, detects sentence boundaries, and formats alphanumeric entities like postal codes and order IDs accurately.<\/p>\n\n\n\n

However, smart transcription is not perfect. The model applies the same foundation-model limitations noted in Google’s documentation, including occasional hallucinations, processing slowness, and rare timeout issues. Production teams should test their own audio before assuming it will work flawlessly.<\/p>\n\n\n\n

Multilingual and Code-Switching Support<\/strong><\/h3>\n\n\n\n

The model automatically detects and transcribes more than 85 languages. It handles regional accents and diverse dialects. Importantly, it supports mid-session language switching, meaning a single audio file containing speakers who alternate between English, Spanish, and Mandarin can be transcribed correctly without manual language specification.<\/p>\n\n\n\n

Custom Vocabulary Biasing<\/strong><\/h3>\n\n\n\n

Specialized terminology matters in transcription. Medical jargon, product names, brand names, and technical acronyms often get mangled by general-purpose transcription models. Gemini 3.5 Transcribe accepts up to 1,000 custom vocabulary terms.<\/p>\n\n\n\n

Google notes that customers typically see the best results with up to around 100 terms. This is because overfitting the vocabulary can sometimes degrade overall transcription accuracy. Custom vocabulary works best for domain-specific words that appear frequently in your audio.<\/p>\n\n\n\n

Speaker Diarization (Recorded Audio Only)<\/strong><\/h3>\n\n\n\n

Gemini 3.5 Transcribe can identify and label individual speakers in pre-recorded audio. The model supports speaker diarization for up to eight speakers. However, attribution for three or more speakers is experimental, meaning the accuracy may be inconsistent.<\/p>\n\n\n\n

The live endpoint does not support speaker diarization at all. This is an important limitation for voice application builders who need real-time speaker identification.<\/p>\n\n\n\n

Word-Level Timestamps (Recorded Audio Only)<\/strong><\/h3>\n\n\n\n

Word-level timestamps mark the exact start and end times for each word in a transcript. This capability is essential for video subtitles, podcast editing, meeting highlight generation, and searchable recordings. It enables downstream applications to link transcript text directly to audio segments.<\/p>\n\n\n\n

Google notes that enabling word-level timestamps can degrade transcription accuracy. Enabling both speaker diarization and word-level timestamps limits recorded audio to 30 minutes per request instead of the standard 1-hour limit. The live endpoint does not support word-level timestamps.<\/p>\n\n\n\n

Pricing Takeaway:<\/strong> The recorded model has an estimated blended rate of about $0.005 per minute, while Transcribe Live is about $0.009 per minute based on Google’s current pricing estimates. Actual billing is token-based.<\/p>\n\n\n\n

Benchmark Performance: What the Numbers Mean<\/strong><\/h2>\n\n\n\n

Artificial Analysis Results<\/strong><\/h3>\n\n\n\n

Google reports that Gemini 3.5 Transcribe achieves an average Word Error Rate (WER) of 4.0% for streaming and 2.6% for non-streaming, as measured by Artificial Analysis. Word Error Rate measures the percentage of words in the output that differ from the reference transcript, calculated as (insertions + deletions + substitutions) divided by total words in the reference.<\/p>\n\n\n\n

Lower WER generally means fewer transcription errors. A 2.6% error rate means roughly 26 errors per 1,000 words, or about one mistake per sentence in typical English speech.<\/p>\n\n\n\n

FLEURS Benchmark Results<\/strong><\/h3>\n\n\n\n

Google also reports performance on the FLEURS (Few-shot Language Evaluation Benchmark) benchmark across a set of top languages and locales:<\/p>\n\n\n\n

Benchmark<\/strong><\/td>Gemini 3.5 Transcribe<\/strong><\/td>Type<\/strong><\/td><\/tr>
FLEURS streaming<\/td>5.50% WER<\/td>Multilingual test set<\/td><\/tr>
FLEURS non-streaming<\/td>5.04% WER<\/td>Multilingual test set<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

The FLEURS benchmark is more conservative than Artificial Analysis and reflects performance across diverse languages and recording conditions.<\/p>\n\n\n\n

Why the Numbers Differ<\/strong><\/h3>\n\n\n\n

The Artificial Analysis and FLEURS benchmarks are fundamentally different evaluation setups. Artificial Analysis uses a specific test methodology optimized for measuring latency and streaming accuracy. FLEURS is a standardized multilingual speech benchmark that includes more diverse languages and recording scenarios.<\/p>\n\n\n\n

These should not be treated as directly interchangeable. The 2.6% WER (Artificial Analysis, non-streaming) and 5.04% WER (FLEURS, non-streaming) are both accurate measurements of Gemini 3.5 Transcribe, but they measure different things using different audio datasets.<\/p>\n\n\n\n

Benchmark Takeaway:<\/strong> Google reports 4.0% streaming WER and 2.6% non-streaming WER in Artificial Analysis, while FLEURS results are 5.50% and 5.04%. Different benchmarks measure different things; don’t compare them directly.<\/p>\n\n\n\n

Beyond WER: What Benchmarks Don’t Measure<\/strong><\/h3>\n\n\n\n

Word Error Rate is useful for comparing models quantitatively, but it doesn’t capture the full user experience. Formatting quality, punctuation, speaker attribution accuracy, latency, handling of domain-specific terminology, code-switching robustness, and self-correction recognition all matter in real-world transcription.<\/p>\n\n\n\n

A model might have a slightly higher WER but produce much more polished transcripts with better speaker labels and punctuation. Similarly, latency matters for live applications, even if accuracy is equivalent.<\/p>\n\n\n\n

Latency Improvement Over Chirp 3<\/strong><\/h3>\n\n\n\n

Google reports a 70% improvement in time to final transcription compared with Chirp 3, based on Artificial Analysis measurements. This means Gemini 3.5 Transcribe produces complete, final transcripts significantly faster than its predecessor, making it more suitable for low-latency voice applications.<\/p>\n\n\n\n

Gemini 3.5 Transcribe Pricing and Cost Examples<\/strong><\/h2>\n\n\n\n

Current API Pricing<\/strong><\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Gemini 3.5 Transcribe pricing follows Google’s token-based model. Billing is calculated by token consumption, not elapsed time, but Google provides minute-based estimates for planning.<\/p>\n\n\n\n

Item<\/strong><\/td>Recorded Audio<\/strong><\/td>Real-Time Streaming<\/strong><\/td><\/tr>
Audio input<\/td>$2.00\/M tokens (~$0.003\/min)<\/td>$3.50\/M tokens (~$0.005\/min)<\/td><\/tr>
Text output<\/td>$12.00\/M tokens (~$0.002\/min)<\/td>$21.00\/M tokens (~$0.004\/min)<\/td><\/tr>
Estimated blended rate<\/td>~$0.005\/min<\/td>~$0.009\/min<\/td><\/tr>
Best for<\/td>Files, meetings, podcasts<\/td>Voice agents, live captions<\/td><\/tr>
Max audio duration<\/td>1 hour (30 min with timestamps\/diarization)<\/td>10 minutes per session<\/td><\/tr>
Word timestamps<\/td>Yes<\/td>No<\/td><\/tr>
Speaker diarization<\/td>Yes<\/td>No<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

These minute-based estimates are Google’s calculations based on assumed token consumption. Your actual costs depend on the exact token usage of your audio and transcription output.<\/p>\n\n\n\n

Practical Cost Examples<\/strong><\/h3>\n\n\n\n

Based on Google’s estimated blended rates:<\/strong><\/p>\n\n\n\n

Gemini 3.5 Transcribe (recorded)<\/strong> at approximately $0.005\/min:<\/p>\n\n\n\n