MyanmarTTS 1.0.0__tar.gz

myanmartts-1.0.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 KdaiP
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

myanmartts-1.0.0/PKG-INFO

@@ -0,0 +1,173 @@
+Metadata-Version: 2.5
+Name: MyanmarTTS
+Version: 1.0.0
+Summary: A from-scratch Burmese (Myanmar) text-to-speech model (CC0).
+Project-URL: Homepage, https://huggingface.co/freococo/MyanmarTTS
+Project-URL: Repository, https://huggingface.co/freococo/MyanmarTTS
+Author: freococo
+License-File: LICENSE
+Keywords: burmese,myanmar,speech,text-to-speech,tts
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: huggingface-hub>=0.20
+Requires-Dist: numba
+Requires-Dist: numpy<2.3
+Requires-Dist: safetensors
+Requires-Dist: scipy
+Requires-Dist: soundfile
+Requires-Dist: torch>=2.0
+Requires-Dist: torchaudio>=2.0
+Requires-Dist: torchdiffeq
+Description-Content-Type: text/markdown
+
+<div align="center">
+
+# StableTTS
+
+Next-generation TTS model using flow-matching and DiT, inspired by [Stable Diffusion 3](https://stability.ai/news/stable-diffusion-3).
+
+
+</div>
+
+## Introduction
+
+As the first open-source TTS model that tried to combine flow-matching and DiT, **StableTTS** is a fast and lightweight TTS model for chinese, english and japanese speech generation. It has 31M parameters. 
+
+✨ **Huggingface demo:** [🤗](https://huggingface.co/spaces/KdaiP/StableTTS1.1)
+
+## News
+
+2024/10: A new autoregressive TTS model is coming soon...
+
+2024/9: 🚀 **StableTTS V1.1 Released** ⭐ Audio quality is largely improved ⭐
+
+⭐ **V1.1 Release Highlights:**
+
+- Fixed critical issues that cause the audio quality being much lower than expected. (Mainly in Mel spectrogram and Attention mask)
+- Introduced U-Net-like long skip connections to the DiT in the Flow-matching Decoder.
+- Use cosine timestep scheduler from [Cosyvoice](https://github.com/FunAudioLLM/CosyVoice)
+- Add support for CFG (Classifier-Free Guidance).
+- Add support for [FireflyGAN vocoder](https://github.com/fishaudio/vocoder/releases/tag/1.0.0).
+- Switched to [torchdiffeq](https://github.com/rtqichen/torchdiffeq) for ODE solvers.
+- Improved Chinese text frontend (partially based on [gpt-sovits2](https://github.com/RVC-Boss/GPT-SoVITS)).
+- Multilingual support (Chinese, English, Japanese) in a single checkpoint.
+- Increased parameters: 10M -> 31M.
+
+
+## Pretrained models
+
+### Text-To-Mel model
+
+Download and place the model in the `./checkpoints` directory, it is ready for inference, finetuning and webui.
+
+| Model Name | Task Details | Dataset | Download Link |
+|:----------:|:------------:|:-------------:|:-------------:|
+| StableTTS | text to mel | 600 hours | [🤗](https://huggingface.co/KdaiP/StableTTS1.1/resolve/main/StableTTS/checkpoint_0.pt)|
+
+### Mel-To-Wav model
+
+Choose a vocoder (`vocos` or `firefly-gan` ) and place it in the `./vocoders/pretrained` directory.
+
+| Model Name | Task Details | Dataset | Download Link |
+|:----------:|:------------:|:-------------:|:-------------:|
+| Vocos | mel to wav | 2k hours | [🤗](https://huggingface.co/KdaiP/StableTTS1.1/resolve/main/vocoders/vocos.pt)|
+| firefly-gan-base | mel to wav | HiFi-16kh | [download from fishaudio](https://github.com/fishaudio/vocoder/releases/download/1.0.0/firefly-gan-base-generator.ckpt)|
+
+## Installation
+
+1. **Install pytorch**: Follow the [official PyTorch guide](https://pytorch.org/get-started/locally/) to install pytorch and torchaudio. We recommend the latest version (tested with PyTorch 2.4 and Python 3.12).
+
+2. **Install Dependencies**: Run the following command to install the required Python packages:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Inference
+
+For detailed inference instructions, please refer to `inference.ipynb`
+
+We also provide a webui based on gradio, please refer to `webui.py`
+
+## Training
+
+StableTTS is designed to be trained easily. We only need text and audio pairs, without any speaker id or extra feature extraction. Here’s how to get started:
+
+### Preparing Your Data
+
+1. **Generate Text and Audio pairs**: Generate the text and audio pair filelist as `./filelists/example.txt`. Some recipes of open-source datasets could be found in `./recipes`.
+
+2. **Run Preprocessing**: Adjust the `DataConfig` in `preprocess.py` to set your input and output paths, then run the script. This will process the audio and text according to your list, outputting a JSON file with paths to mel features and phonemes. 
+
+**Note: Process multilingual data separately by changing the `language` setting in `DataConfig`**
+
+### Start training
+
+1. **Adjust Training Configuration**:  In `config.py`, modify `TrainConfig` to set your file list path and adjust training parameters (such as batch_size) as needed.
+
+2. **Start the Training Process**: Launch `train.py` to start training your model. 
+
+Note: For finetuning, download the pretrained model and place it in the `model_save_path` directory specified in  `TrainConfig`. Training script will automatically detect and load the pretrained checkpoint.
+
+### (Optional) Vocoder training
+
+The `./vocoder/vocos` folder contains the training and finetuning codes for vocos vocoder.
+
+For other types of vocoders, we recommend to train by using [fishaudio vocoder](https://github.com/fishaudio/vocoder): an uniform interface for developing various vocoders. We use the same spectrogram transform so the vocoders trained is compatible with StableTTS.
+
+## Model structure
+
+<div align="center">
+
+<p style="text-align: center;">
+  <img src="./figures/structure.jpg" height="512"/>
+</p>
+
+</div>
+
+- We use the Diffusion Convolution Transformer block from [Hierspeech++](https://github.com/sh-lee-prml/HierSpeechpp), which is a combination of original [DiT](https://github.com/sh-lee-prml/HierSpeechpp) and [FFT](https://arxiv.org/pdf/1905.09263.pdf)(Feed forward Transformer from fastspeech) for better prosody.
+
+- In flow-matching decoder, we add a [FiLM layer](https://arxiv.org/abs/1709.07871) before DiT block to condition timestep embedding into model.
+
+## References
+
+The development of our models heavily relies on insights and code from various projects. We express our heartfelt thanks to the creators of the following:
+
+### Direct Inspirations
+
+[Matcha TTS](https://github.com/shivammehta25/Matcha-TTS): Essential flow-matching code.
+
+[Grad TTS](https://github.com/huawei-noah/Speech-Backbones/tree/main/Grad-TTS): Diffusion model structure.
+
+[Stable Diffusion 3](https://stability.ai/news/stable-diffusion-3): Idea of combining flow-matching and DiT.
+
+[Vits](https://github.com/jaywalnut310/vits): Code style and MAS insights, DistributedBucketSampler.
+
+### Additional References:
+
+[plowtts-pytorch](https://github.com/p0p4k/pflowtts_pytorch): codes of MAS in training
+
+[Bert-VITS2](https://github.com/Plachtaa/VITS-fast-fine-tuning) : numba version of MAS and modern pytorch codes of Vits
+
+[fish-speech](https://github.com/fishaudio/fish-speech): dataclass usage and mel-spectrogram transforms using torchaudio, gradio webui
+
+[gpt-sovits](https://github.com/RVC-Boss/GPT-SoVITS): melstyle encoder for voice clone
+
+[coqui xtts](https://huggingface.co/spaces/coqui/xtts): gradio webui
+
+Chinese Dirtionary Of DiffSinger: [Multi-langs_Dictionary](https://github.com/colstone/Multi-langs_Dictionary) and [atonyxu's fork](https://github.com/atonyxu/Multi-langs_Dictionary)
+
+## TODO
+
+- [x] Release pretrained models.
+- [x] Support Japanese language.
+- [x] User friendly preprocess and inference script.
+- [x] Enhance documentation and citations.
+- [x] Release multilingual checkpoint.
+
+## Disclaimer
+
+Any organization or individual is prohibited from using any technology in this repo to generate or edit someone's speech without his/her consent, including but not limited to government leaders, political figures, and celebrities. If you do not comply with this item, you could be in violation of copyright laws.

myanmartts-1.0.0/README.md

@@ -0,0 +1,148 @@
+<div align="center">
+
+# StableTTS
+
+Next-generation TTS model using flow-matching and DiT, inspired by [Stable Diffusion 3](https://stability.ai/news/stable-diffusion-3).
+
+
+</div>
+
+## Introduction
+
+As the first open-source TTS model that tried to combine flow-matching and DiT, **StableTTS** is a fast and lightweight TTS model for chinese, english and japanese speech generation. It has 31M parameters. 
+
+✨ **Huggingface demo:** [🤗](https://huggingface.co/spaces/KdaiP/StableTTS1.1)
+
+## News
+
+2024/10: A new autoregressive TTS model is coming soon...
+
+2024/9: 🚀 **StableTTS V1.1 Released** ⭐ Audio quality is largely improved ⭐
+
+⭐ **V1.1 Release Highlights:**
+
+- Fixed critical issues that cause the audio quality being much lower than expected. (Mainly in Mel spectrogram and Attention mask)
+- Introduced U-Net-like long skip connections to the DiT in the Flow-matching Decoder.
+- Use cosine timestep scheduler from [Cosyvoice](https://github.com/FunAudioLLM/CosyVoice)
+- Add support for CFG (Classifier-Free Guidance).
+- Add support for [FireflyGAN vocoder](https://github.com/fishaudio/vocoder/releases/tag/1.0.0).
+- Switched to [torchdiffeq](https://github.com/rtqichen/torchdiffeq) for ODE solvers.
+- Improved Chinese text frontend (partially based on [gpt-sovits2](https://github.com/RVC-Boss/GPT-SoVITS)).
+- Multilingual support (Chinese, English, Japanese) in a single checkpoint.
+- Increased parameters: 10M -> 31M.
+
+
+## Pretrained models
+
+### Text-To-Mel model
+
+Download and place the model in the `./checkpoints` directory, it is ready for inference, finetuning and webui.
+
+| Model Name | Task Details | Dataset | Download Link |
+|:----------:|:------------:|:-------------:|:-------------:|
+| StableTTS | text to mel | 600 hours | [🤗](https://huggingface.co/KdaiP/StableTTS1.1/resolve/main/StableTTS/checkpoint_0.pt)|
+
+### Mel-To-Wav model
+
+Choose a vocoder (`vocos` or `firefly-gan` ) and place it in the `./vocoders/pretrained` directory.
+
+| Model Name | Task Details | Dataset | Download Link |
+|:----------:|:------------:|:-------------:|:-------------:|
+| Vocos | mel to wav | 2k hours | [🤗](https://huggingface.co/KdaiP/StableTTS1.1/resolve/main/vocoders/vocos.pt)|
+| firefly-gan-base | mel to wav | HiFi-16kh | [download from fishaudio](https://github.com/fishaudio/vocoder/releases/download/1.0.0/firefly-gan-base-generator.ckpt)|
+
+## Installation
+
+1. **Install pytorch**: Follow the [official PyTorch guide](https://pytorch.org/get-started/locally/) to install pytorch and torchaudio. We recommend the latest version (tested with PyTorch 2.4 and Python 3.12).
+
+2. **Install Dependencies**: Run the following command to install the required Python packages:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Inference
+
+For detailed inference instructions, please refer to `inference.ipynb`
+
+We also provide a webui based on gradio, please refer to `webui.py`
+
+## Training
+
+StableTTS is designed to be trained easily. We only need text and audio pairs, without any speaker id or extra feature extraction. Here’s how to get started:
+
+### Preparing Your Data
+
+1. **Generate Text and Audio pairs**: Generate the text and audio pair filelist as `./filelists/example.txt`. Some recipes of open-source datasets could be found in `./recipes`.
+
+2. **Run Preprocessing**: Adjust the `DataConfig` in `preprocess.py` to set your input and output paths, then run the script. This will process the audio and text according to your list, outputting a JSON file with paths to mel features and phonemes. 
+
+**Note: Process multilingual data separately by changing the `language` setting in `DataConfig`**
+
+### Start training
+
+1. **Adjust Training Configuration**:  In `config.py`, modify `TrainConfig` to set your file list path and adjust training parameters (such as batch_size) as needed.
+
+2. **Start the Training Process**: Launch `train.py` to start training your model. 
+
+Note: For finetuning, download the pretrained model and place it in the `model_save_path` directory specified in  `TrainConfig`. Training script will automatically detect and load the pretrained checkpoint.
+
+### (Optional) Vocoder training
+
+The `./vocoder/vocos` folder contains the training and finetuning codes for vocos vocoder.
+
+For other types of vocoders, we recommend to train by using [fishaudio vocoder](https://github.com/fishaudio/vocoder): an uniform interface for developing various vocoders. We use the same spectrogram transform so the vocoders trained is compatible with StableTTS.
+
+## Model structure
+
+<div align="center">
+
+<p style="text-align: center;">
+  <img src="./figures/structure.jpg" height="512"/>
+</p>
+
+</div>
+
+- We use the Diffusion Convolution Transformer block from [Hierspeech++](https://github.com/sh-lee-prml/HierSpeechpp), which is a combination of original [DiT](https://github.com/sh-lee-prml/HierSpeechpp) and [FFT](https://arxiv.org/pdf/1905.09263.pdf)(Feed forward Transformer from fastspeech) for better prosody.
+
+- In flow-matching decoder, we add a [FiLM layer](https://arxiv.org/abs/1709.07871) before DiT block to condition timestep embedding into model.
+
+## References
+
+The development of our models heavily relies on insights and code from various projects. We express our heartfelt thanks to the creators of the following:
+
+### Direct Inspirations
+
+[Matcha TTS](https://github.com/shivammehta25/Matcha-TTS): Essential flow-matching code.
+
+[Grad TTS](https://github.com/huawei-noah/Speech-Backbones/tree/main/Grad-TTS): Diffusion model structure.
+
+[Stable Diffusion 3](https://stability.ai/news/stable-diffusion-3): Idea of combining flow-matching and DiT.
+
+[Vits](https://github.com/jaywalnut310/vits): Code style and MAS insights, DistributedBucketSampler.
+
+### Additional References:
+
+[plowtts-pytorch](https://github.com/p0p4k/pflowtts_pytorch): codes of MAS in training
+
+[Bert-VITS2](https://github.com/Plachtaa/VITS-fast-fine-tuning) : numba version of MAS and modern pytorch codes of Vits
+
+[fish-speech](https://github.com/fishaudio/fish-speech): dataclass usage and mel-spectrogram transforms using torchaudio, gradio webui
+
+[gpt-sovits](https://github.com/RVC-Boss/GPT-SoVITS): melstyle encoder for voice clone
+
+[coqui xtts](https://huggingface.co/spaces/coqui/xtts): gradio webui
+
+Chinese Dirtionary Of DiffSinger: [Multi-langs_Dictionary](https://github.com/colstone/Multi-langs_Dictionary) and [atonyxu's fork](https://github.com/atonyxu/Multi-langs_Dictionary)
+
+## TODO
+
+- [x] Release pretrained models.
+- [x] Support Japanese language.
+- [x] User friendly preprocess and inference script.
+- [x] Enhance documentation and citations.
+- [x] Release multilingual checkpoint.
+
+## Disclaimer
+
+Any organization or individual is prohibited from using any technology in this repo to generate or edit someone's speech without his/her consent, including but not limited to government leaders, political figures, and celebrities. If you do not comply with this item, you could be in violation of copyright laws.

myanmartts-1.0.0/pyproject.toml

@@ -0,0 +1,36 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "MyanmarTTS"
+version = "1.0.0"
+description = "A from-scratch Burmese (Myanmar) text-to-speech model (CC0)."
+readme = "README.md"
+requires-python = ">=3.9"
+authors = [{ name = "freococo" }]
+keywords = ["tts", "burmese", "myanmar", "speech", "text-to-speech"]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Operating System :: OS Independent",
+    "Topic :: Multimedia :: Sound/Audio :: Speech",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "torch>=2.0",
+    "torchaudio>=2.0",
+    "huggingface_hub>=0.20",
+    "numpy<2.3",
+    "numba",
+    "scipy",
+    "soundfile",
+    "safetensors",
+    "torchdiffeq",
+]
+
+[project.urls]
+Homepage = "https://huggingface.co/freococo/MyanmarTTS"
+Repository = "https://huggingface.co/freococo/MyanmarTTS"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/myanmar_tts"]

myanmartts-1.0.0/src/myanmar_tts/__init__.py

@@ -0,0 +1,42 @@
+"""MyanmarTTS - Burmese text-to-speech, from scratch, CC0."""
+import os
+from huggingface_hub import hf_hub_download
+
+HF_REPO = "freococo/MyanmarTTS"
+MODEL_FILE = "model_fp16.safetensors"
+VOCODER_FILE = "vocos.pt"
+REF_AUDIO = "samples/sample_0.wav"
+
+__version__ = "1.0.0"
+
+
+class MyanmarTTS:
+    def __init__(self, device="cuda"):
+        self.device = device
+        self._model = None
+        self._ref = None
+
+    def _load(self):
+        if self._model is not None:
+            return
+        print(f"Downloading model files from {HF_REPO}...")
+        model_path = hf_hub_download(repo_id=HF_REPO, filename=MODEL_FILE)
+        vocos_path = hf_hub_download(repo_id=HF_REPO, filename=VOCODER_FILE)
+        self._ref = hf_hub_download(repo_id=HF_REPO, filename=REF_AUDIO)
+
+        from .api import StableTTSAPI
+        print("Loading model...")
+        self._model = StableTTSAPI(model_path, vocos_path, "vocos")
+        self._model.to(self.device)
+        print("Ready.")
+
+    def tts(self, text):
+        """Return a numpy float32 waveform at 44100 Hz."""
+        self._load()
+        audio, _ = self._model.inference(
+            text, self._ref, "burmese", step=32, solver="dopri5", cfg=3.0
+        )
+        return audio.squeeze(0).cpu().numpy()
+
+
+__all__ = ["MyanmarTTS"]

myanmartts-1.0.0/src/myanmar_tts/api.py

@@ -0,0 +1,82 @@
+import torch
+import torch.nn as nn
+from dataclasses import asdict
+
+from .utils.audio import LogMelSpectrogram
+from .config import ModelConfig, MelConfig
+from .models.model import StableTTS
+
+from .text import symbols
+from .text import cleaned_text_to_sequence
+from .text.burmese import burmese_to_ipa2
+from .datas.dataset import intersperse
+from .utils.audio import load_and_resample_audio
+
+
+def get_vocoder(model_path, model_name='vocos'):
+    if model_name == 'vocos':
+        from .vocoders.vocos.models.model import Vocos
+        from .config import VocosConfig, MelConfig
+        vocoder = Vocos(VocosConfig(), MelConfig())
+        vocoder.load_state_dict(torch.load(model_path, weights_only=True, map_location='cpu'))
+        vocoder.eval()
+    else:
+        raise NotImplementedError(f"Unsupported vocoder: {model_name}")
+    return vocoder
+
+
+class StableTTSAPI(nn.Module):
+    def __init__(self, tts_model_path, vocoder_model_path, vocoder_name='vocos'):
+        super().__init__()
+        self.mel_config = MelConfig()
+        self.tts_model_config = ModelConfig()
+
+        self.mel_extractor = LogMelSpectrogram(**asdict(self.mel_config))
+
+        self.tts_model = StableTTS(len(symbols), self.mel_config.n_mels, **asdict(self.tts_model_config))
+
+        if tts_model_path.endswith(".safetensors"):
+            from safetensors.torch import load_file
+            state = load_file(tts_model_path)
+        else:
+            state = torch.load(tts_model_path, map_location='cpu', weights_only=True)
+        self.tts_model.load_state_dict(state)
+        self.tts_model.eval()
+
+        self.vocoder_model = get_vocoder(vocoder_model_path, vocoder_name)
+        self.vocoder_model.eval()
+
+        self.g2p_mapping = {
+            'burmese': burmese_to_ipa2,
+        }
+        self.supported_languages = self.g2p_mapping.keys()
+
+    @torch.inference_mode()
+    def inference(self, text, ref_audio, language, step, temperature=1.0,
+                  length_scale=1.0, solver=None, cfg=3.0):
+        device = next(self.parameters()).device
+        phonemizer = self.g2p_mapping.get(language)
+        if phonemizer is None:
+            raise ValueError(f"Unsupported language: {language}")
+
+        text = phonemizer(text)
+        text = torch.tensor(
+            intersperse(cleaned_text_to_sequence(text), item=0),
+            dtype=torch.long, device=device
+        ).unsqueeze(0)
+        text_length = torch.tensor([text.size(-1)], dtype=torch.long, device=device)
+
+        ref_audio = load_and_resample_audio(ref_audio, self.mel_config.sample_rate).to(device)
+        ref_audio = self.mel_extractor(ref_audio)
+
+        mel_output = self.tts_model.synthesise(
+            text, text_length, step, temperature, ref_audio,
+            length_scale, solver, cfg
+        )['decoder_outputs']
+        audio_output = self.vocoder_model(mel_output)
+        return audio_output.cpu(), mel_output.cpu()
+
+    def get_params(self):
+        tts_param = sum(p.numel() for p in self.tts_model.parameters()) / 1e6
+        vocoder_param = sum(p.numel() for p in self.vocoder_model.parameters()) / 1e6
+        return tts_param, vocoder_param

myanmartts-1.0.0/src/myanmar_tts/burmese.py

@@ -0,0 +1,42 @@
+"""
+Character-level text frontend for Burmese.
+No real G2P: we clean the text and pass characters through directly.
+Matches the interface of english_to_ipa2 / japanese_to_ipa2 (returns a list).
+"""
+import re
+
+# Burmese Unicode ranges
+_BURMESE_RANGES = (
+    (0x1000, 0x109F),   # Myanmar
+    (0xAA60, 0xAA7F),   # Myanmar Extended-A
+    (0xA9E0, 0xA9FF),   # Myanmar Extended-B
+)
+
+# Allow these through the cleaner; everything else gets dropped
+_ALLOWED_ASCII_PUNCT = set(".,!?'\"-:;()")
+_ALLOWED_BURMESE_PUNCT = set("\u104A\u104B\u104C\u104D\u104E\u104F")
+
+
+def _is_burmese(ch):
+    o = ord(ch)
+    return any(lo <= o <= hi for lo, hi in _BURMESE_RANGES)
+
+
+def clean_burmese(text):
+    """Keep only Burmese chars, spaces, and whitelisted punctuation."""
+    out = []
+    for ch in text:
+        if ch in (" ", "\t", "\n", "\r"):
+            out.append(" ")
+        elif _is_burmese(ch):
+            out.append(ch)
+        elif ch in _ALLOWED_ASCII_PUNCT or ch in _ALLOWED_BURMESE_PUNCT:
+            out.append(ch)
+        # else: drop silently
+    return re.sub(r"\s+", " ", "".join(out)).strip()
+
+
+def burmese_to_ipa2(text):
+    """Return a list of character tokens (matches the IPA-G2P interface)."""
+    cleaned = clean_burmese(text)
+    return list(cleaned)

myanmartts-1.0.0/src/myanmar_tts/config.py

@@ -0,0 +1,50 @@
+from dataclasses import dataclass
+
+@dataclass
+class MelConfig:
+    sample_rate: int = 44100
+    n_fft: int = 2048
+    win_length: int = 2048
+    hop_length: int = 512
+    f_min: float = 0.0
+    f_max: float = None
+    pad: int = 0
+    n_mels: int = 128
+    center: bool = False
+    pad_mode: str = "reflect"
+    mel_scale: str = "slaney"
+    
+    def __post_init__(self):
+        if self.pad == 0:
+            self.pad = (self.n_fft - self.hop_length) // 2
+            
+@dataclass
+class ModelConfig:
+    hidden_channels: int = 256
+    filter_channels: int = 1024
+    n_heads: int = 4
+    n_enc_layers: int = 3 
+    n_dec_layers: int = 6 
+    kernel_size: int = 3
+    p_dropout: int = 0.1
+    gin_channels: int = 256
+            
+@dataclass
+class TrainConfig:
+    train_dataset_path: str = 'filelists/filelist.json'
+    test_dataset_path: str = 'filelists/filelist.json' # not used
+    batch_size: int = 32
+    learning_rate: float = 1e-4
+    num_epochs: int = 50
+    model_save_path: str = './checkpoints'
+    log_dir: str = './runs'
+    log_interval: int = 5
+    save_interval: int = 1
+    warmup_steps: int = 200
+    
+@dataclass
+class VocosConfig:
+    input_channels: int = 128
+    dim: int = 512
+    intermediate_dim: int = 1536
+    num_layers: int = 8