Supported Languages

VoxCPMTTS uses openbmb/VoxCPM2. The model officially supports 30 languages and does not require a language tag for synthesis: provide text in a supported language directly. The language field in the VoxCPMTTS UI and API is a compatibility hint used by related features such as transcription and timestamp alignment; VoxCPM2 detects the synthesis language from the input text.

Source: official VoxCPM2 model card

Official Language List

All 30 official languages are listed individually below. The benchmark manifest contains prompts for every language. The published speech-quality run measured 24 languages because its Qwen3-ASR judge did not advertise the other six.

# Language Common code Published benchmark
1 Arabic ar Measured
2 Burmese my Not judged by the configured ASR
3 Chinese zh Measured
4 Danish da Measured
5 Dutch nl Measured
6 English en Measured
7 Finnish fi Measured
8 French fr Measured
9 German de Measured
10 Greek el Measured
11 Hebrew he Not judged by the configured ASR
12 Hindi hi Measured
13 Indonesian id Measured
14 Italian it Measured
15 Japanese ja Measured
16 Khmer km Not judged by the configured ASR
17 Korean ko Measured
18 Lao lo Not judged by the configured ASR
19 Malay ms Measured
20 Norwegian no, nb, nn Not judged by the configured ASR
21 Polish pl Measured
22 Portuguese pt Measured
23 Russian ru Measured
24 Spanish es Measured
25 Swahili sw Not judged by the configured ASR
26 Swedish sv Measured
27 Tagalog / Filipino tl, fil Measured as Filipino
28 Thai th Measured
29 Turkish tr Measured
30 Vietnamese vi Measured

Chinese Dialects

The model card additionally documents these Chinese dialect groups:

  • Sichuanese (四川话)
  • Cantonese (粤语)
  • Wu Chinese (吴语)
  • Northeastern Mandarin (东北话)
  • Henan dialect (河南话)
  • Shaanxi dialect (陕西话)
  • Shandong dialect (山东话)
  • Tianjin dialect (天津话)
  • Southern Min (闽南话)

Dialect generation is driven by the input text and optional voice direction. VoxCPMTTS does not expose these dialects as separate model checkpoints.

Practical Notes

  • Official support means the language was included in the model’s documented coverage. Quality can still vary with the text, voice design, reference recording, punctuation, and generation settings.
  • A reference recording should be clean and should match its transcript exactly when transcript-guided cloning is used.
  • Long-form protection is enabled by default in VoxCPMTTS. It divides long passages at natural boundaries to reduce accumulated ringing or background artifacts; users can disable it in Generation settings or with "protect_long_audio": false in the native API.
  • The reproducible speech-quality benchmark covers the intersection of VoxCPM2 and the configured Qwen3-ASR judge. See benchmarks/speech-quality/DETAILS.md for measured results rather than treating official support as a quality ranking.