Speak a non-English or mixed-language text without drifting language
Four or more companies' docs show a full fit.
What counts as done
The requested language is supported.
Measured, for our own tests: a native listener accepts the pronunciation and accent, numbers and dates are read in the requested language, and a 20-sentence passage never switches language or script mid-sentence. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: read this in Japanese
, say this Hinglish message
, speak the German part in German and the English part in English
Does the job 7
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Cartesia Sonic TTS API + MCP B Cartesia AI, Inc. | Sonic 3.6 speaks 44 languages; language or locale is required on every request and sets normalisation, with romanised Hindi handled, and clones are built per language. source |
| ElevenLabs Text to Speech API + MCP BB Eleven Labs Inc. | language_code enforces the language for the model and normalisation on v4, v3 and Flash (not Multilingual v2), v4 covers 90-plus languages, and acc-39 runs a Spanish podcast on it. source |
| Google Cloud Text-to-Speech B Google LLC | Each request names a voice languageCode and locale-bound voices cover many languages; CJK is billed per character, not per byte; acc-14 keeps Gemini as default over Kokoro for its language and voice range. source |
| HeyGen Voice BB HeyGen Technology, Inc. | language takes a lowercase ISO 639-1 code from 41 supported languages and the speech follows it as given; when omitted HeyGen detects it and every response names the language used in Content-Language. source |
| Inworld Realtime TTS-2 C Theai, Inc. dba Inworld AI | A language parameter and inline <lang> tags per span pin each part of a mixed text to a BCP-47 code on every endpoint, across 623 language and locale codes in three tiers. source |
| Murf TTS API + MCP BB Murf Inc. | locale on every request pins the language of a multilingual voice (replacing multiNativeLocale), each voice lists its supported locales, and Falcon 2.1 is described as switching languages within a sentence. source |
| Soniox Text-to-Speech BB Soniox Inc. | A required language field sets delivery and accent, embedded foreign words are read in place without splitting, every voice speaks all 60-plus languages, and script limits for Serbian, Bosnian and Kazakh are stated. source |
Does the job, with a limit 15
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Amazon Polly BB Amazon Web Services, Inc. | Voices are bound to one locale, so a single-language text cannot drift, but mixed text needs lang tags and only a few voices (Arjun, Kiara en-IN and hi-IN) are bilingual; generative voices cover 24 locales. source |
| Audio8-TTS-Realistic 0.6B D Edge0 (edge0.ai), which names no legal entity or location on its site. The technical report's authors give their affiliation as AutoArk, and commit metadata in Edge0-AI/Audio8_TTS uses an autoark.ai domain | 23 languages, but no language parameter, and the card's own table shows 21 to 49 percent word error in Arabic, Hindi, Finnish, Vietnamese and Thai. source |
| Azure AI Speech text-to-speech BB Microsoft Corporation | Locale-bound voices for 100-plus locales and a lang element on multilingual voices pin the language, but acc-27 shows a multilingual voice drifting into Chinese on digits with no root cause found. source |
| Deepgram Text-to-Speech (Aura-2, Flux TTS) A Deepgram, Inc. | Flux voices come in 42 languages and the language is fixed by the model string, so no drift, but normalisation runs in only ten languages and the others are Early Access with numbers written out. source |
| Fish Audio Drama 3 C 39 AI, Inc. (the licensor named in the Fish Audio Community License). fish.audio's terms of use name Hanabi AI Inc. as the provider of the hosted services | The card lists about 80 languages and Fish Audio documents language tags for a phrase, but those tags are experimental and the model may switch language on multilingual text, and acc-31 reports failures in Russian and German as well as French. source |
| Fish Audio S2 Pro D 39 AI, Inc. (the licensor named in the Fish Audio Research License). fish.audio's terms of use name Hanabi AI Inc. as the provider of the hosted services | 80-plus languages in three tiers, but the local server has no language parameter, so nothing pins the reading on mixed or digit-heavy text. source |
| Fish Audio TTS API B Hanabi AI Inc. | 83 languages on s2.1-pro, but the language tag only steers normalisation and timestamps and is not passed to the model, the docs say to leave it unset for mixed text and that drama-3 may switch language on its own, and CJK text bills at about three bytes a character. source |
| Ming-3.0-tts-0.5B C inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group) | language_id pins Chinese or English and is required for German, Spanish, French, Korean and Japanese, with Sichuanese and Cantonese dialects, but only seven languages and the card lists Korean accuracy and multilingual speaker similarity as gaps. source |
| Ming-3.0-tts-tiny C inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI) | language_id or target_language pins one of seven languages plus two dialects, but the card says Korean accuracy and multilingual speaker similarity remain gaps. source |
| MiniMax Speech (T2A) C Nanonoble Pte. Ltd. | language_boost takes one of about 40 language names or auto and the async guide lists 40 languages, but the field is described as enhancing recognition of minority languages and dialects rather than locking the reading, so drift on mixed text is not ruled out. source |
| OmniVoice C k2-fsa, the Next-gen Kaldi open-source project. The paper's authors are at Xiaomi Corp.; no company is named as the licensor | 646 languages and a language_id field in the batch JSONL, but cross-lingual cloning carries the reference accent, voice design is trained on Chinese and English only, and nothing is said about mixed-script drift. source |
| Qwen3-TTS 12Hz 1.7B CustomVoice C Alibaba Cloud (the copyright holder in the Qwen licences; RDAP names Alibaba Cloud (Singapore) Private Limited as registrant of qwen.ai, from the first batch's check) | A language argument pins one of ten languages or Auto and the package rejects unsupported ones, but acc-24 found Japanese output slipping into Mandarin mid-sentence, which the parameter did not prevent. source |
| Rime TTS API + MCP C Rime Labs, Inc. | A BCP-47 language parameter pins the reading and the Coda WebSocket lists supported tags in ready.languages, but the cloud API covers ten languages, 185 of 313 Coda voices are English, and a lexicon is refused in Arabic, Dutch, Hindi and Japanese. source |
| Smallest.ai Lightning TTS B Smallest, Inc. | A required language code pins the voice's language and Pro covers 31 languages with auto routing, but a mismatched voice and model pairing is not rejected and can return the wrong voice or no audio, and nothing documents digit reading. source |
| VieNeu-TTS v3 Turbo C Phạm Nguyễn Ngọc Bảo, an individual (named on the card as author; Hugging Face account pnnbao-ump, GitHub account pnnbao97). No company found | English and Vietnamese code-switching is a stated feature, but no other language is spoken and there is no language parameter. source |
Couldn't check 1
The docs were closed to us, or silent. These carry no mark either way.
| Listing | Why |
|---|---|
| Resemble AI Text-to-Speech API D Resemble AI, Inc. | No language or locale parameter and no language list appear on the synthesis pages read, and no account names Resemble. |
What stops agents finishing
- the voice drifting into another language with no language parameter to stop it (acc-24 Qwen3-TTS into Mandarin, acc-27 Azure multilingual voice into Chinese)
- English-only models (Deepgram Flux) or only two languages (VieNeu)
- a per-character price that triples for CJK text billed by UTF-8 byte (Fish Audio) or counts each Chinese character as two (Azure)
Around it
Before: translation.translate-text, inference.chat-completion. After: text-to-speech.voice-agent-turn, text-to-speech.long-form-narration.
Where this comes from
Weight 6 of 100 in Text-to-speech: 7 of 46 accounts tagged (acc-08, 14, 18, 23, 24, 27, 33); the drift failures (acc-24, 27) are the sharpest evidence, but only seven accounts needed a language other than English
The same as data: /api/v1/use-cases/text-to-speech.other-language-correct.json, or the MCP tool get_use_case with id text-to-speech.other-language-correct. Other jobs in this category: /api/v1/niches?category=text-to-speech.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.