Sauti TTS V2 (Chatterbox Swahili LoRA)
This is a Swahili text-to-speech (TTS) model fine-tuned on top of the English Chatterbox Multilingual backbone using a LoRA adapter. It was developed by Msingi-AI as part of the Sauti TTS V2 project.
Model Details
- Base Model: Chatterbox Multilingual (English backbone + multilingual grapheme tokenizer)
- Training Method: LoRA fine-tuning
- Language: Swahili (
sw) - License: MIT
Training Data
The model was fine-tuned on Google's WaxalNLP swa_tts studio dataset (CC-BY-4.0).
The dataset was filtered using a transcript-agreement check, dropping clips where an Automatic Speech Recognition (ASR) judge detected a Character Error Rate (CER) > 10% against the reference text.
Evaluation Results
The model was evaluated against a held-out evaluation set of 48 Swahili sentences across various stress categories (general, code-switching, named entities, numbers/dates).
Intelligibility is measured by Word Error Rate (WER) and Character Error Rate (CER) under two ASR judges: zero-shot openai/whisper-large-v3 and a Swahili-fine-tuned Whisper (Jacaranda-Health/ASR-STT). Swahili's agglutinative morphology tends to inflate WER, making CER the headline metric. UTMOS is used as a proxy for audio quality.
| Scope | openai/whisper-large-v3 WER / CER | Jacaranda-Health/ASR-STT WER / CER | UTMOS |
|---|---|---|---|
| Overall | 0.388 / 0.088 | 0.139 / 0.038 | 3.887 |
| Overall (Plain Swahili) | 0.384 / 0.086 | 0.092 / 0.022 | 3.871 |
| Code Switch | 0.400 / 0.094 | 0.279 / 0.084 | 3.934 |
| General | 0.320 / 0.069 | 0.148 / 0.036 | 3.872 |
| Named Entities | 0.421 / 0.094 | 0.027 / 0.005 | 3.897 |
| Numbers & Dates | 0.411 / 0.095 | 0.100 / 0.026 | 3.844 |
Usage
The weights provided here are the LoRA adapter (new_lang_adapter directory). They must be loaded on top of the original Chatterbox backbone using the Chatterbox fine-tuning toolkit.
- Downloads last month
- -