effusiveperiscope / xtts-repl

Quick REPL for XTTS

What is this

Quick and dirty REPL for running XTTS

Installation

pip install -r requirements.txt Install desired models and reference audios in models and ref_audio respectively.

Directory structure for added models

models/
    Celestia/
        config.json
        model_can_name_anything_as_long_as_it_is.pth
ref_audio/
    reference_audio1.flac
    reference_audio2.wav
    suggested_to_name_audios_after_speaker.wav

Usage

python main.py

Command: h
        q                    Exits the program
        h                    Show this help
        M                    Select model checkpoint and load model
        R                    Select reference audios
        refr                 Refresh files
        temp <temperature>   Set temperature
        T <text>               TTS
        t                    Repeat last TTS request
        lang                 Select language
        s                    Show state (temp, lang, ref audio)
        p <idx>              Preview audio by index
        pre                  Select audio for preview
        S <indices>          Save audio by index
        Save                 Select audio to save
Command: T Hemanth stepped forward, only to be intercepted by one of Orgon's body guards.
Generating texts: 100%|██████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:05<00:00,  2.50s/it]
Command: p 0
Command: s 0

Pruning

python prune.py ckpt_to_prune.pth can prune weights not used for inference from the model, reducing disk size and file I/O time. This will change hashes for precomputed latents.