ONNX runtime still imports torch/torchaudio in `onnx_tts_runtime.py`

Summary

The README states:

No PyTorch dependency during inference: it runs directly on ONNX Runtime CPU.

However, onnx_tts_runtime.py (current main branch, line 12-13) still has:

import torch
import torchaudio

And these are actually used in the _load_reference_audio method (around line 443-451):

waveform, sample_rate = torchaudio.load(...)
waveform = waveform.to(torch.float32)
waveform = torchaudio.functional.resample(waveform, sample_rate, target_sample_rate)

Impact

• Anyone trying to run the ONNX-only deployment without PyTorch installed will get an ImportError when app_onnx.py imports from onnx_tts_runtime.
• This defeats the purpose of the lightweight ONNX deployment path, especially on resource-constrained devices (e.g., ARM boards, mobile) where installing PyTorch is impractical.

Suggested Fix

Replace the torch/torchaudio usage in _load_reference_audio with lightweight alternatives:

• Audio loading: soundfile.read() or librosa.load()
• Float32 conversion: NumPy astype(np.float32)
• Resampling: scipy.signal.resample or librosa.resample

This would make the ONNX deployment truly PyTorch-free.

Environment

• Discovered while deploying on Android (proot Ubuntu, ARM64) where PyTorch cannot be installed.
• Repository cloned from main branch as of 2026-06-04.

Note: This issue was drafted with the assistance of an AI assistant (Operit). The findings are based on direct source code inspection.

Start with onnx_tts_runtime.py, especially the imports and _load_reference_audio around lines 12–13 and 443–451; then inspect how app_onnx.py imports it. Verify the ONNX path can load and resample reference audio without torch or torchaudio, and confirm the documented PyTorch-free inference path works on the relevant deployment setup.