Google's new voice models list Malay, and price calls by the minute

Voice is the interface most Malaysian businesses actually run on. Customers call, voice notes pile up, and staff take orders while doing three other things. On 15 September 2026 Google put new voice models behind that problem, with a published per-minute price.[1][2]

The release covers Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking for speech-to-speech work, and Gemini 3.5 Transcribe for speech-to-text. Nothing in the announcement mentions Malaysia.[1]

The announcement names three models. Gemini 3.8 Live is a native speech-to-speech model that Google says can perform tasks while maintaining a conversation. Gemini 3.8 Live Extended Thinking adds configurable background reasoning for longer requests. Gemini 3.5 Transcribe, released a month earlier, is the dedicated speech-to-text model. All are reachable through the Gemini API and Google AI Studio.[1]

Google lists capabilities rather than customers. The Live models support asynchronous function calling, so a tool call runs in the background while audio keeps streaming, plus visual context and what Google calls alphanumeric precision for confirmation codes and claim numbers. Seven integration partners are named for media streaming, a fair measure of the plumbing between a model and a service.[1]

Google quotes USD0.005 a minute for audio input and USD0.018 a minute for audio output on the Live models, marking the output figure as an estimate based on USD12 per million tokens. The published pricing page repeats both numbers and shows the token rates behind them: USD3.00 per million audio input tokens and USD12.00 per million audio output tokens on the paid tier.[1][2]

Transcribe is cheaper. The pricing page puts Gemini 3.5 Transcribe at about USD0.003 a minute for audio input and USD0.002 a minute for text output, an effective blended rate Google estimates at roughly USD0.005 a minute. The streaming version is estimated at about USD0.009 a minute. A free tier exists for all of them, on the stated condition that free-tier data improves Google's products.[2]

The Live API guide states that the interface supports 99 languages, carrying Malay as ms, Tamil as ta and Indonesian as id. The Gemini 3.5 Transcribe table is different: more than 85 locales, with Malay as ms-MY, Cantonese and Mandarin listed separately, and no Tamil entry at all. Google's blog uses a third figure, 97 or more languages.[3][4][1]

Two documented features matter more than the count. Transcribe handles code-switching inside and between sentences without manual configuration, which is how most Malaysians actually speak, and it accepts a custom vocabulary list of up to 1,000 terms to bias recognition towards company names and product codes. The published word error rates are averages across all supported languages, not Malay results.[4][1]