| --- |
| language: [ha, yo, ig, en] |
| license: mit |
| pipeline_tag: text-to-speech |
| tags: [text-to-speech, voice-cloning, chatterbox, multilingual, hausa, yoruba, igbo, nigerian-english, african-languages] |
| base_model: ResembleAI/chatterbox |
| --- |
| |
| # Naija TTS — multilingual Hausa / Yoruba / Igbo / Nigerian English |
|
|
| One Chatterbox model covering four Nigerian languages, with zero-shot voice |
| cloning. LoRA fine-tune of Chatterbox Multilingual (MIT), with the grapheme |
| vocabulary extended for the characters these languages need. |
|
|
| Base Chatterbox ships 23 languages, of which Swahili is the only African one. |
| This model adds Hausa, Yoruba, Igbo and Nigerian English — the last under its |
| own `en-ng` tag so base English is left intact. |
|
|
| ## Evaluation |
|
|
| | Language | n | CER | CER floor | WER | Speaker SIM | RTF | |
| |---|---|---|---|---|---|---| |
| | Hausa (`ha`) | 40 | 8.83 % | 4.87 % | 32.35 % | n/a | 0.680 | |
| | Yoruba (`yo`) | 40 | 45.02 % | 12.13 % | 77.73 % | n/a | 1.065 | |
| | Igbo (`ig`) | 40 | 14.87 % | 8.41 % | 42.57 % | n/a | 0.840 | |
| | Nigerian English (`en-ng`) | 40 | 12.27 % | 13.09 % | 31.44 % | n/a | 0.932 | |
|
|
| **Read each CER against its own floor**, not against zero. The floor is what |
| MMS-ASR scores on genuine recordings of that language — its own error rate. |
| A CER at the floor means the synthesized speech is as intelligible as real |
| speech. The floors differ substantially between languages; a raw CER comparison |
| across languages is meaningless. |
|
|
| ## Usage |
|
|
| ```python |
| import torchaudio as ta |
| import chatterbox.mtl_tts as M |
| from chatterbox.mtl_tts import ChatterboxMultilingualTTS |
| |
| for code, name in [("ha","Hausa"), ("yo","Yoruba"), ("ig","Igbo"), |
| ("en-ng","Nigerian English")]: |
| M.SUPPORTED_LANGUAGES[code] = name |
| |
| model = ChatterboxMultilingualTTS.from_local("path/to/repo", device="cuda") |
| wav = model.generate("Sannu da zuwa.", language_id="ha", |
| audio_prompt_path="reference.wav") |
| ta.save("out.wav", wav, model.sr) |
| ``` |
|
|
| Normalize text to Unicode NFC and keep the languages' special characters as |
| written — for Yoruba this includes the combining tone marks. |
|
|
| ## Limitations |
|
|
| - **Hausa tone** is not modelled: Boko orthography marks neither tone nor vowel |
| length, so a character-input model has no signal for it. Yoruba and Igbo write |
| tone with diacritics and are not subject to this. |
| - Long utterances can truncate (Chatterbox forces EOS on long tails). |
| - The corpus was capped per language for balance, so no language sees its full |
| available data. |
| - Outputs carry Resemble's Perth watermark, inherited from the base model. |
|
|
| ## Licence |
| MIT, from base Chatterbox. |
|
|
