AI4Bharat TTS Model Inference for Tibetan (Continuing Post-Training)

[tenzinchoedon]

Link to previous card: #2 (comment)

Description

This phase focuses on running inference on the fine-tuned AI4Bharat Indic-TTS model. The objective is to generate high-quality, natural-sounding Tibetan speech and validate the model's performance. Necessary debugging and adjustments will also be carried out if issues arise during inference.

Completion Criteria

• Generate Tibetan speech audio files from textual inputs.
• Resolve any issues related to phoneme mismatches or inference configurations.

Implementation

1. Inference Execution

• Run inference using the fine-tuned Tibetan TTS model and prepared Tibetan text input.
• Save generated audio files for evaluation.

2. Debugging

• Investigate and resolve issues like unsupported phoneme languages (bo) or incorrect configurations.
• Make adjustments to phoneme language or preprocessing as required.

3. Evaluation

• Evaluate the quality of the generated speech.
• Compare the generated outputs with the expected results to spot any differences.

Subtasks

• Validate inference setup and configurations.
• Run inference on sample Tibetan text input.
• Save and analyze generated audio outputs for debugging and quality assessment.
• Document issues, if any, and implement required adjustments.

Card Reviewer

• @gangagyatso4364

[tenzinchoedon]

Wav files from training on Weights & Biases (WandB):

TrainAudio
EvalAudio

[tenzinchoedon]

Link to the documentation: https://docs.google.com/document/d/1pj5ZWiiUJ0lLKvaBC76ITcAtmGYLGklbknXH6KCE5Oo/edit?usp=sharing