piper-gpl-revamped-voices

This is a experimental repository containing currently 1 model operating at 32-KHz (Effective 16-KHz) sample rate.

[Overview]

This project's intention is to create both higher quality Piper-TTS (Vits family) models, and partially show effects or lack there of, during distillation from another voice mode.

Due to the training dataset being generated by another diffusion-based TTS model, and because of the inherent architecture of VITS, it's expected to encounter, out of ordinary quantization artifacts or misalignment in pronunciations during inferencing.

Currently available voices are:

• en_US-mei-high-onnx

There is a checkpoint available for continuous training:

• epoch=699-step=200120.ckpt

[Voice Samples]

--Source vs Inference - Comparisons--

Source Audio 01

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/source_01.wav"></audio>

Inference Audio 01

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/inference_01.wav"></audio>

Source Audio 02

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/source_02.wav"></audio>

Inference Audio 02

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/inference_02.wav"></audio>

Free Inferencing

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/free_inference_01.wav"></audio>

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/free_inference_02.wav"></audio>

<audio controls src="https://huggingface.co/VINUK/piper-gpl-revamped-voices/resolve/main/demo_audio/free_inference_03.wav"></audio>

[Inference Guide]

Inferencing is really simple. Follow the steps below.

This is tested working within a Python==3.14 conda environment.

NOTE: You may change the noise_scale and noise_w to adjust variations for your workload. Or reference to PIPER_REPO. Furthermore, the best settings found are noise_scale : 0.75 and noise_w_scale : 0.75

1. Install the piper-tts package using pip.

pip install piper-tts

2. Download the model file from above, alongside the corresponding .json file.

3. Put them both into the same folder.

4. Create a .py with your desired name and add the below code.

import wave
from piper import PiperVoice

voice = PiperVoice.load("/path/to/YOUR_DOWNLOADED_MODEL.onnx")
with wave.open("test.wav", "wb") as wav_file:
    voice.synthesize_wav("Welcome to the world of speech synthesis!", wav_file)

Remember to change YOUR_DOWNLOAD_MODEL.onnxto the actual one.

[Training Details]

This model is trained on combination of CMU_ARCTIC and VCTK utterances, with a total dataset size of 14920 samples, at 16-bit 32KHz on a RTX6000 96GB for 700 epochs with varying batch sizes.

Training Args Used

python -m piper.train fit \
--data.voice_name "mei" \
--data.csv_path /teamspace/studios/this_studio/meta_data.csv \
--data.audio_dir /teamspace/studios/this_studio/32k_sampling \
--data.cache_dir /teamspace/studios/this_studio/compiled_data \
--data.config_path /teamspace/studios/this_studio/onnx_models/config.json \
--data.espeak_voice "en-us" \
--data.batch_size 100 \
--data.num_workers 10 \
--data.trim_silence False \
--model.sample_rate 32000 \
--model.num_speakers 1 \
--model.hidden_channels 256 \
--model.inter_channels 256 \
--model.filter_channels 1024 \
--model.n_heads 4 \
--model.n_layers 6 \
--model.filter_length 1024 \
--model.hop_length 256 \
--model.win_length 1024 \
--model.mel_channels 100 \
--model.n_layers_q 3 \
--model.p_dropout 0.05 \
--model.gin_channels 0 \
--model.use_sdp True \
--model.resblock 1 \
--model.segment_size 20480 \
--model.resblock_kernel_sizes "[3, 7, 11]" \
--model.resblock_dilation_sizes "[[1, 3, 5], [1, 3, 5], [1, 3, 5]]" \
--model.upsample_rates "[8, 8, 2, 2]" \
--model.upsample_initial_channel 512 \
--model.upsample_kernel_sizes "[16, 16, 4, 4]" \
--model.learning_rate 7e-5 \
--model.lr_decay 1.0 \
--model.learning_rate_d 6e-5 \
--model.lr_decay_d 1.0 \
--model.c_kl 1.0 \
--model.c_mel 50 \
--trainer.default_root_dir /teamspace/studios/this_studio/ \
--trainer.max_epochs 700 \
--trainer.log_every_n_steps 10