Gemini 3.5 Transcribe lists Thai among 85 plus languages at about 0.18 baht a minute

Google has opened Gemini 3.5 Transcribe in public preview through the Gemini API and Google AI Studio, and Thai (th-TH) appears in the official Supported Languages list alongside more than 85 other languages. Marketing Oops reported the release on 27 August 2026, including paid-tier pricing of roughly $0.005 per minute for file processing, which converts to approximately 0.18 baht a minute at the exchange rate implied in that report.

For anyone in Thailand who has ever paid a person to type up an hour of interview audio, that number is the story. The rest of this article is what sits underneath it: which modes exist, what they change about the text you get back, which features quietly do not work in streaming mode yet, and how to read the speed benchmark that is already being repeated without its caveat.

What Google put into public preview

Marketing Oops reported on 27 August 2026 that Gemini 3.5 Transcribe is available in public preview through two front doors: the Gemini API for anyone building it into software, and Google AI Studio for anyone who wants to paste in a file and look at the output. Google says the model succeeds Chirp 3, its previous speech-to-text model.

There are two things being shipped, not one. The first is the file-processing model, which takes an audio file and returns a transcript. The second is a separate streaming model, gemini-3.5-transcribe-live, which runs over WebSocket and is built for low-latency live captioning and voice interfaces. They are not the same product with a different switch flipped, and as the section on feature gaps below explains, they do not currently have the same feature set.

The supported-language coverage is the part that matters for this market. Thai is not an unofficial side effect of a multilingual model here. The language code th-TH is on the official Supported Languages list, which is a different level of commitment from a model that happens to produce passable Thai when you feed it Thai.

Verbatim and Smart Transcription do two different jobs

Marketing Oops described two transcription modes, and choosing the wrong one will quietly damage your work.

Verbatim keeps what was actually said. Filler words stay in. Repetitions stay in. Mid-sentence self-corrections stay in. That is the mode for interviews and research, where the hesitation before an answer is data and where changing a respondent's words changes the finding.

Smart Transcription cleans as it writes. It strips fillers such as "เอ่อ" and "อืม", resolves repetitions and self-corrections, and auto-formats the output into paragraphs, bullets, dates, numbers and currency. The worked example given in the report is a speaker saying "Tuesday, no wait, Wednesday, 2pm" and the system rendering the final text as a Wednesday 14:00 appointment.

Read that example again, because it is doing more than tidying. The model resolved a correction and normalised a time format, which is exactly what you want from meeting notes and exactly what you do not want from a research transcript that someone may later quote. Smart Transcription produces text that is closer to publishable. Verbatim produces text that is closer to evidence. Here is how the two compare on the points the report described.

Speaker labels, timestamps and a 1,000-term custom vocabulary

Beyond the two modes, Marketing Oops listed four capabilities. Speaker diarization separates up to 8 speakers. Word-level timestamps mark when each word was spoken. A custom vocabulary accepts up to 1,000 terms, which the report describes as product names, company names, abbreviations and industry jargon. Code-switching is handled with automatic detection across the 85 or more supported languages.

The custom vocabulary limit is the one to plan around. A thousand terms is enough to load an entire Thai product catalogue's naming conventions, the staff roster of a call centre, and the acronym soup that any B2B company speaks in. Speech models fail most visibly on proper nouns, and proper nouns are precisely what a brand cares about seeing spelled correctly in a transcript.

Code-switching deserves its own line for a Thai audience. Thai business conversation moves between Thai and English inside a single sentence as a matter of routine, not as an exception. A model that detects the switch automatically, rather than requiring you to declare one language per file, is describing the way people here actually talk. The report did not publish accuracy figures for Thai-English code-switching specifically, so treat that capability as claimed rather than measured.

Two features do not work in Live mode yet

This is the detail most likely to break somebody's plan, and Marketing Oops stated it plainly: speaker diarization and word-level timestamps currently work on file processing and are not yet supported in the streaming Live mode.

If your idea was a live meeting assistant that labels who is speaking in real time, that idea does not work today with this model. If your idea was live captions with no speaker labels, that works. The practical pattern the split suggests is to caption live for the room and then run the recording through file processing afterwards for the labelled, timestamped version.

The table below sets the two paths against each other, including the prices reported by Marketing Oops. The baht figures are approximate conversions of the dollar prices, not prices Google quoted in baht.

Marketing Oops also noted that a free tier is available. The report did not spell out the free tier's quota or limits, so anyone planning a pilot should check that inside the Gemini API documentation rather than assume a volume.

How to read the 70 percent speed number

Marketing Oops cites testing by Artificial Analysis showing that Time to Final Transcription improved by roughly 70 percent versus the predecessor model. The article then does something that gets lost the moment a number travels: it explicitly cautions that this does not mean transcription is 70 percent faster in every situation. What improved is the wait for a completed transcript, and that wait drops significantly.

The distinction is worth holding onto. Time to Final Transcription is one metric on one test set. It is not a promise about your audio, your file lengths, your network, your language mix or your concurrency. A team that budgets a workflow on "70 percent faster" and then measures something else will be disappointed by a model that did exactly what was claimed.

The report did not publish the test methodology behind that figure, the audio conditions, the sample size or the language distribution. Absent those, the honest reading is directional: the wait got substantially shorter on a third-party benchmark, and your own numbers are the ones that decide whether the workflow is viable.

What cheap, accurate transcription actually unlocks for a marketing team

Nothing in the Marketing Oops report concerns SEO or marketing use. Google announced a speech-to-text model, not a content strategy. What follows is reasoning from the published price and features, and should be read that way.

At roughly 0.18 baht a minute for file processing, the cost of turning spoken audio into text stops being a budget line and starts being a rounding error. A hundred hours of recorded sales calls is 6,000 minutes, which at the reported rate works out to about $30, or roughly 1,080 baht, using the same approximate conversion. That arithmetic is derived from the published per-minute price, not a figure Google quoted.

Once the price of transcription collapses, a lot of audio that was previously write-only becomes searchable text. Webinars, podcasts, recorded product training, conference panels, YouTube video and call-centre recordings all become documents. Those documents are raw material for content marketing that is grounded in things your customers and your own experts actually said, rather than in a keyword brief.

Three specific uses follow from the feature list rather than from ambition. Word-level timestamps are what caption files are made of, which matters for YouTube assets where captions carry the message for viewers watching without sound. Speaker diarization turns a two-hour panel into attributable quotes. And the questions customers ask out loud on sales calls, once they exist as text, are the most honest source of FAQ content a company has, because nobody invented them.

The last implication is about answer engines. Text that is self-contained, question-shaped and specific is the kind of passage that gets lifted into an AI answer. Transcripts are full of that shape naturally, since a spoken answer to a spoken question is already a complete unit. Cleaning them into published pages is a plausible route to the citations that generative engine optimisation chases. Google did not say this. It follows from what the model does.

What this means for Thai marketers

The immediate change is that Thai audio has an official, priced, documented path into text through a major model, at a rate that makes bulk processing defensible. Thai has often sat in the second tier of automatic speech recognition support, present in a model's capability but absent from its official list. Being on the Supported Languages list is a statement about what Google intends to keep working.

The realistic first project is not a live product. It is a back catalogue. Most Thai companies of any age are sitting on recorded webinars, training sessions and call archives that nobody has ever read. File processing is where the features are complete and the price is lowest, so that is where the value is easiest to prove before anyone builds a live feature.

Two cautions belong next to that. First, this is a public preview, and Marketing Oops did not report a general availability date. Preview terms and quotas can change. Second, the report published no Thai-specific accuracy figures at all. The language is officially supported; how well it performs on your accents, your industry jargon and your recording quality is an empirical question that a short pilot on real files answers better than any launch article.

What the report did not say

Clear reporting includes the gaps. Marketing Oops did not publish a general availability date, a word error rate for Thai or any other language, quota or rate limits for the paid or free tiers, data residency or retention terms, or a shutdown date for Chirp 3. It also made no claim about how the model should be used for marketing or search. Anything you read elsewhere asserting those specifics is not coming from this report.

Common questions (FAQ)

Is Thai really supported, or does the model just happen to handle it?

Thai is officially supported, listed as th-TH on the Supported Languages list among more than 85 languages. Marketing Oops reported that on 27 August 2026. What the report did not include is any Thai-specific accuracy measurement, so official support tells you Google commits to the language, not how the output will read on your particular recordings.

How much does it cost to transcribe a large archive?

On the paid tier, file processing is about $0.005 per minute, approximately 0.18 baht. Live streaming is about $0.009 per minute, approximately 0.32 baht. The baht figures are approximate conversions of the dollar prices rather than prices Google quoted in baht, and a free tier is available although the report did not describe its limits.

Can I use it for live captioning at an event?

Yes, through gemini-3.5-transcribe-live, a separate streaming model that runs over WebSocket for low-latency captioning and voice interfaces. The limitation to plan around is that speaker diarization and word-level timestamps are not yet supported in Live mode, so real-time output will not tell you who is speaking.

Does it really transcribe 70 percent faster?

No, and the source says so directly. Artificial Analysis testing cited by Marketing Oops shows Time to Final Transcription improved roughly 70 percent versus the predecessor, and the article explicitly cautions that this does not mean transcription is 70 percent faster in every situation. The improvement is in how long you wait for a finished transcript.

Do I need to change anything today?

No. This is a public preview of an optional model, and Marketing Oops reported no deadline, migration requirement or general availability date. Google says Gemini 3.5 Transcribe succeeds Chirp 3, but the report gave no date on which Chirp 3 stops working, so existing pipelines are not on a published clock.

If you are weighing whether an archive of Thai audio is worth turning into text, the useful next step is small: run a handful of real files, in both modes, and read what comes back before you plan anything around it. Relevant Audience is happy to talk through what that content is worth once it exists.

Source: Marketing Oops.