跳到正文
DEV Community· Chandra Shekhar·· 3 天前AI 評分65

離開英語後,文字轉語音的價格如何變化

The language tax: what text-to-speech costs once you leave English

AI 導讀

作者整理 Voicedex 收錄的 297 個文字轉語音(TTS)模型及 126 家供應商公開價格,發現不同語言的最低標價差距明顯。西班牙語、法語、德語、印地語、阿拉伯語、日語和中文的最低價為每 1 million 個字符 $4,斯瓦希里語為 $15,約魯巴語為 $40;這些均是供應商公佈的最低級別價格。

正文

Cover image for The language tax: what text-to-speech costs once you leave English

Chandra Shekhar

I track prices for 297 text-to-speech models across 126 providers in an open catalog, Voicedex. Prices are normalized to US dollars per 1 million characters and taken only from the vendors' own pricing pages. When I sorted them by language, a pattern jumped out.

The cheapest listed TTS price, by language (pay-as-you-go prices; discounts that need a monthly plan are left out):

Language TTS models listing it Priced Cheapest listed
Spanish, French, German, Hindi, Arabic, Japanese, Chinese 103–139 47–63 $4
Bengali, Tamil, Vietnamese 67–85 28–44 $4
Swahili 35 13 $15
Amharic 17 6 $15
Zulu 13 7 $15
Hausa 12 5 $15
Yoruba 9 2 $40

For the world's largest languages, competition pushes the floor to $4 per million characters (the "standard" voice tiers from the big clouds). For Swahili, spoken by well over 100 million people, the floor is almost four times higher. For Yoruba it is ten times higher, and only two models publish a price at all.

Speech-to-text doesn't have the same tax. The cheapest listed STT for Swahili is $0.10 per hour (Soniox; Azure's promotional rate through 31 December matches it), the same as for Hindi, Arabic or Tamil, and Yoruba starts at $0.13. A likely reason: large multilingual speech-recognition models cover many languages at once, while synthesis still depends more on voices built language by language.

The overall spread is huge in both directions:

  • TTS: about $1 to $250 per 1M characters, median $30 (94 of 297 models publish a comparable price).
  • STT: $0.03 to $4.68 per hour, median $0.72 (223 of 444).
  • Speech-to-speech: $0.78 to $60 per hour, median $3 (13 of 74).

Most models don't publish a comparable per-unit price at all: they're credit plans, enterprise-only or per-token. That's a pricing problem of its own.

Models disappear fast. 36 speech models currently have an announced shutdown date, 26 of them within the next 90 days. Thirteen providers in the catalog have shut down or stopped selling entirely, including PlayHT, Coqui, LMNT and Resemble AI.

Caveats:

  • "Cheapest listed" is the lowest published tier. The $4 tiers are older-generation voices; newer neural voices cost more, and quality differs a lot.
  • Language support is what each provider claims. A model listing a language doesn't guarantee it sounds native.
  • Prices move. These are from 18 November 2026; the live numbers are at voicedex.dev, and every change is in the git history.

If you build voice products for African or South Asian markets, the comparison pages are the fastest way to see your real options: Swahili TTS, Hindi TTS, STT by language.

Corrections welcome: if a price here is wrong, open an issue with the vendor's link and it gets fixed in the open.

來源:DEV Community · dev.to