"We want to set up a multilingual call center for inbound tourists, but securing operators is impossible," "We can build an English IVR, but Spanish, Tagalog, and Vietnamese are out of reach"—along with the rebound in inbound tourism, inquiries like these have noticeably surged since February.
The answer to this demand may lie in Gemini 3.1 Flash TTS, released by Google on April 16. Supporting over 70 languages, it enables natural voice generation with emotional expression at low latency and low cost. In this article, we organize implementation patterns and cost estimates when incorporating it into custom development.
Three features of Flash TTS
Compared to traditional cloud TTS services, Flash TTS stands out in the following three areas.
- Support for 70 languages: Broadly covers Japanese, English, and Chinese, as well as major Asian languages such as Tagalog, Swahili, Thai, Vietnamese, and Hindi
- Control over emotional expression: Speaking styles can be specified using
excited,calm, andapologetic, producing voices tailored to brand tone - Low latency: Consistently delivers time-to-first-audio under 300 ms, making it viable for real-time phone responses
This marks the point where the evolution of the Gemini family, which we tracked in the trend of the Gemini app replacing Google Assistant, has finally expanded into the voice infrastructure layer.
Application scenarios: Ready models for client development
Scenario 1: Multilingual IVR (interactive voice response)
A setup that takes incoming telephone inquiries from inbound tourists, automatically detects the language, and returns response audio via Flash TTS.
[1] 着信 → 「Please press 1 for English, 2 for 日本語, ...」
[2] 音声認識で言語判定（Gemini 3.1 の ASR）
[3] FAQ 検索 → 該当言語で Flash TTS 生成 → 再生
[4] 解決しない場合のみ、英語 or 日本語オペレーターへエスカレーションBecause it can reduce operator workload by 60% to 80%, 24-hour coverage can be assembled at a realistic cost.
Scenario 2: Audio guides for tourist destinations
A configuration where visitors scan a QR code to open a web app and instantly generate commentary on the cultural property they are viewing in their own language. This eliminates the need to prepare pre-recorded audio files in multiple languages.
Scenario 3: Voice commerce (smart speakers and in-vehicle systems)
A user experience that completes product searches, recommendations, and purchase confirmations entirely through conversational audio. With Flash TTS's emotional expression, you can switch between a bright tone for recommending products and a calm tone for confirmation messages.
Scenario 4: Multilingual announcements for hospitals and public facilities
Waiting number announcements automatically switch based on the language setting configured at registration. This is an area that local governments and medical institutions cannot solve through staffing alone.
Implementation patterns: The battle against audio latency
To create an experience that feels seamless even during telephone calls, a latency budget must be established from the design phase.
| Process | Target latency | Method |
|---|---|---|
| Speech recognition | 200ms | Gemini 3.1 ASR + streaming |
| Intent understanding + answer generation | 500ms | Caching + simple intent classification |
| TTS generation | 300ms | Flash TTS streamed output |
| Total | Within 1,000 ms | Feels natural for a phone call |
In particular, designing the TTS side to stream sequentially from the beginning of each sentence cuts perceived wait time in half.
Brand tone control
Explicitly specify emotion, speed, and speaker gender via prompts.
{
"model": "gemini-3.1-flash-tts",
"voice": {
"name": "ja-JP-Standard-Female-01",
"style": "calm_professional",
"pace": "slow"
},
"text": "お電話ありがとうございます。GleamHub サポートです。..."
}Unnatural delivery, such as sounding overly cheerful or mechanical, can be improved by 90% simply through combinations of style and pace.
Cost estimate: IVR at a scale of 100,000 calls per month
Assuming 100,000 calls per month (averaging 2 minutes across 5 languages) as an alternative to an inbound customer call center.
| Item | Monthly |
|---|---|
| ASR (speech recognition) 200,000 minutes | Approx. 400,000 JPY |
| LLM inference (Gemini 3.1) | Approx. 300,000 JPY |
| TTS (Flash TTS) output audio 200,000 minutes | Approx. 600,000 JPY |
| Twilio call charges | Approx. 800,000 JPY |
| Total | Approx. 2,100,000 JPY/month |
In contrast, human operators for 5 languages across 3 shifts run 9,000,000 to 12,000,000 JPY per month. The reduction is 7,000,000 to 9,900,000 JPY per month, representing a cost transformation approaching nearly 100 million JPY annually.
Three pitfalls during implementation
- Handling dialects and accents: Dialects like Kansai-ben or Okinawan are generated in standard Japanese, so compensate in prompts when dialect support is required
- Pronunciation of proper nouns: Brand names and place names will falter unless locked down using pronunciation dictionaries
- Legal regulations: Certain countries mandate advance disclosure when using AI voices, requiring verification during global rollouts
The principle of "designing to clearly disclose that it is an AI," which we wrote about in our article on AI agent transparency and UX, is equally critical for voice.
Summary — An era where language barriers no longer require hiring solutions
In areas where multilingual support cannot be solved by manual labor alone, such as tourist sites, healthcare, government, and retail, Flash TTS serves as a structural solution. Shifting the mindset away from adding operators toward redesigning primary responses with AI voice is accelerating on the ground in 2026.
If you are looking to build an inbound call center or turn multilingual tourist guidance into an app, please feel free to reach out via our contact form.
