index-tts-lora: Fine-Tune Bilibili index-tts for Speech Synthesis

Overview

index-tts-lora is a Python fine-tuning project built on Bilibili's index-tts. It adds LoRA workflows for adapting the speech synthesis model to single- or multi-speaker audio, aiming to improve prosody and naturalness.

The project is for developers and speech teams who want to adapt index-tts using their own recorded speech and transcripts. The README demonstrates a Chinese speaker dataset and includes Chinese, English, and mixed-language synthesis examples; it does not describe a general-purpose voice service or a turnkey hosted product.

Key Features

• LoRA fine-tuning workflows for Bilibili's index-tts.
• Supports both single-speaker and multi-speaker setups.
• Extracts audio tokens and speaker conditions from audio and transcript lists.
• Generates processed training and validation metadata, plus speaker condition information.
• Provides training and inference entry points.
• Includes example outputs for Chinese, English, and mixed Chinese-English text.

Use Cases

• A speech researcher adapting index-tts to a specific speaker's recordings.
• A developer testing whether a small collection of speaker audio can improve the target voice's prosody.
• A team building a multi-speaker synthesis prototype on top of index-tts.
• A practitioner evaluating synthesis for Chinese, English, or mixed-language text using the supplied examples as a reference.

Getting Started

Prepare an audio list with each audio path and transcript separated by a tab, then extract tokens and speaker conditions:

python tools/extract_codec.py --audio_list ${audio_list} --extract_condition
python train.py
python indextts/infer.py

See the README for the data format and workflow details.

Alternatives

• Spark-TTS: Spark-TTS offers zero-shot voice cloning and controllable generation, while index-tts-lora focuses on LoRA fine-tuning of IndexTTS for speaker adaptation.
• ChatTTS: ChatTTS targets conversational speech with pause and laughter controls, while index-tts-lora adds LoRA fine-tuning for IndexTTS speaker adaptation.
• chatterbox: Chatterbox provides multilingual synthesis and voice cloning, while index-tts-lora specializes in LoRA fine-tuning of IndexTTS for single or multiple speakers.
• VoxCPM: VoxCPM focuses on tokenizer-free multilingual synthesis, voice design, and cloning, while index-tts-lora fine-tunes IndexTTS with LoRA for speaker adaptation.

Considerations

• Project health is marked Quiet: there have been no commits in the last 90 days and no release in the past year. No issues were closed in that period, so there is little recent issue-handling activity to assess.
• Work is concentrated among two contributors, with one responsible for most contributions. This may make maintenance and review capacity a concern.
• The README does not specify hardware requirements, provide a deployment guide, or document a complete dependency and environment setup. Confirm what the index-tts training and inference stack requires before planning a run.
• The repository reports tests, but no CI configuration is present in the supplied project facts.
• The project is licensed under Apache-2.0; review its terms when redistributing or incorporating the code.