EvoTalk

A single-speaker text-to-speech acoustic model built from scratch in PyTorch. A standalone architecture that predicts mel spectrograms from phonemes, vocoded natively with Vocos at 24 kHz.

Links

• GitHub: https://github.com/akkii2006/EvoTalk
• Hugging Face: https://huggingface.co/akkiisfrommars/EvoTalk

Status

The current v1 model has overfitted to its training data (Hi-Fi TTS speaker 9017), so generalization to new speakers or unseen prosody is limited. A second version addressing this is in active development. The model does however produce intelligible audio output for single-speaker synthesis.

Architecture

CosmicFish-like: transformer encoder and decoder (GQA, RoPE, SwiGLU, RMSNorm) around a variance adaptor that predicts per-phoneme duration, pitch, and energy, then expands the enriched sequence to frame level with a length regulator. ~80M parameters.

Training Data

• Hi-Fi TTS (MikhailT/hifi-tts), speaker 9017 (male), ~53 hours, ~51k utterances.
• Text to ARPABET via espeak, durations via MMS forced alignment.
• Log-magnitude mel (100 bands, f_max 12 kHz) matched exactly to the Vocos vocoder config.
• Continuous pitch contour and per-frame energy extracted on GPU.

Pipeline

Stage	Script	Purpose
Prepare	prepare.py	G2P, alignment, feature extraction, dataset + metadata
Train	train.py	Main training with masked L1/MSE losses, AMP, cosine LR
Finetune	finetune.py / long_finetune.py	Long-form and extra-long utterance stages
Synthesize	inference.py	Interactive REPL with sentence/clause chunking
Predict	predict.py	Render the predicted mel spectrogram as an image
Sweep	sweep.py	Render one script across every checkpoint

Usage

python prepare.py --out_dir data
python train.py --data_dir data
python inference.py

Install dependencies with pip install -r requirements.txt.

Graphs

Training curves from graphs/.

Prediction

Predicted mel spectrogram from python predict.py --text "Hello! How are you? I am EvoTalk a TTS model built by Mistyoz AI." --out mel.png.

License

Apache 2.0