athmanemokraoui/kabyle-tts

Kabyle TTS # Kabyle TTS - ONNX Inference with Correct Tokenization A production-ready Kabyle (Taqbaylit) text-to-speech solution using the `facebook/mms-tts-kab` ONNX model with correct MMS tokenization. ## Overview This script works around critical bugs in the `phoonnx` library (all versions including 1.3.4a1) that prevent proper Kabyle speech synthesis: - **Bug 1:** `VoiceConfig.from_dict()` overwrites `config["blank"] = "_"` instead of using `"|"` from the config - **Bug 2:** `TTSTokenizer.from_phoonnx_config()` ignores config fields (`add_blank_char`, `blank_between`, etc.) - **Bug 3:** The phoneme map doesn't include `"_"`, so `blank_id` becomes `None` - **Bug 4:** Without valid `blank_id`, `intersperse_blank_char()` does nothing - **Bug 5:** Result: tokens without blanks → gibberish audio ## Solution This implementation: 1. Uses the **original MMS tokenizer** (`transformers.VitsTokenizer`) for correct tokenization 2. Runs **ONNX inference directly** via `onnxruntime` 3. Manually implements blank token insertion matching MMS VITS behavior 4. Maps missing characters (p, v, o) to closest Kabyle equivalents ## Requirements ```bash pip install transformers torch onnxruntime-gpu huggingface_hub scipy ``` For CPU-only: ```bash pip install transformers torch onnxruntime huggingface_hub scipy ``` ## Quick Start ### In Python ```python from kabyle_tts_suppress_mapped import KabyleTTS # Initialize tts = KabyleTTS(use_cuda=True) # Synthesize tts.speak("Azul, amek i telliḍ ?", "output.wav") # With speed control (0.8 = slower, 1.2 = faster) tts.speak("Taqbaylit d tutlayt tamezwarut n Lezzayer.", speed=0.8) ``` ### In Google Colab Copy the entire script into a single cell and run. All dependencies will be installed automatically. ## Character Mapping The MMS model was trained on Kabyle text without certain characters. The following mappings are applied: | Character | Maps To | Reason | |---|---|---| | `p` | `b` | 'p' not in Kabyle alphabet | | `v` | `f` | 'v' not in …