Hangry LabsUseful tools, easy to run

OmniVoiceTTS examples

Listen to voice examples or SSML-H conversations.

Voices and languages

Choose one of 20 speech languages to hear ten voice-variety samples, randomized introductions and a cross-language clone of the shared English reference voice. The original player supports playback, seeking, mute and a shared volume control. Selecting a language also translates the gallery. Norwegian has translated gallery controls with English recordings. Arabic and Urdu use right-to-left gallery content while the website header keeps its own direction.

The six website locales serve the actual translated READMEs. Other speech languages remain selectable within the gallery; direct links such as ?lang=ar still work. The language query parameter remembers a speech selection while lang selects the website documentation language.

SSML-H conversations

The conversation showcase includes a featured two-character exchange and two additional dialogues, their transcripts, voice declarations, request settings and copyable SSML-H source. These are generated examples, not live inference.

Local preview

Run task pages for the complete website, task pages-build for a static build, or task pages-test for staging and routing checks. These commands do not load speech models or use a GPU. See the website guide.

The index.md and ssml-h.md wrappers include the existing standalone HTML content in the shared website shell. Player code, translations, manifests and recordings remain shared between both presentations. Opening index.html or ssml-h.html directly still works when the assets are kept together. To preview the standalone files over HTTP:

python -m http.server 8055

Open http://localhost:8055/examples/.

Recordings and provenance

Generated language showcase recordings live under assets/<language>/:

  • random/: ten native-language culture/humor samples. Plain-text samples rotate voice-design profiles. Samples with bracket expression tags such as [laughter], [sigh], and [surprise-ah] intentionally use no voice-design instruct because OmniVoice can produce unstable non-speech audio when voice design is combined with bracket tags.
  • intro/: three translated OmniVoiceTTS/Hangry Labs introduction candidates
  • clone/: one reference-clone candidate using examples/original_clone.mp3
  • assets/manifest.json: exact text, language, and file metadata for the generated samples

The gallery selects these recordings from the shared manifest; no model is loaded for playback.

This Hangry Labs fork is intentionally focused on local inference, browser UI, HTTP API, and baked Docker images. Upstream training, fine-tuning, data preparation, and benchmark-evaluation examples were removed from this fork; use the original k2-fsa/OmniVoice repository for those workflows.