How We Fixed IndexTTS2 on Apple Silicon (The torchaudio.save Bug)

We spent hours debugging IndexTTS2 on a Mac Mini M4. Every single output sounded like static — choppy, distorted, unusable. We blamed BigVGAN. We blamed MPS. We tried CPU mode, different voice samples, different diffusion steps. Nothing worked.

The model was producing perfect audio the entire time. The bug was in one line of the save pipeline.

The Symptoms

If you're running IndexTTS2 on Apple Silicon (or honestly, any non-CUDA setup) and getting audio that sounds like:

• Static or white noise
• Heavily distorted/clipped speech
• "Choppy" output that's technically speech but sounds terrible
• Waveforms that look like solid blocks instead of audio

You probably have this bug.

The Root Cause

In indextts/infer_v2.py, the audio save pipeline does this:

# The original code (BROKEN)
wav = torch.clamp(32767 * wav, -32767.0, 32767.0)
torchaudio.save(output_path, wav.type(torch.int16), sampling_rate)

This looks reasonable. BigVGAN outputs float audio in [-1, 1], you scale it to int16 range, cast to int16, and save. Right?

Here's the proof:

# Saving int16 with torchaudio — it rescales!
>>> wav = torch.tensor([[0.0, 5000.0, -5000.0, 20000.0, -20000.0]])
>>> wav_int16 = wav.type(torch.int16)
>>> torchaudio.save('/tmp/test.wav', wav_int16, 22050)
>>> sf.read('/tmp/test.wav', dtype='int16')
array([0, 32767, -32768, 32767, -32768])

Every value that isn't zero gets slammed to ±32767. This is torchaudio "helpfully" normalizing your audio.

The Fix

# The fix — save as normalized float with explicit bit depth
torchaudio.save(
    output_path,
    (wav / 32767.0).clamp(-1.0, 1.0).float(),
    sampling_rate,
    bits_per_sample=16
)

That's it. Same file format, same quality, but torchaudio now knows the data is already properly scaled.

Two More Fixes You'll Need

1. Missing Qwen Emotion Model

IndexTTS2's emotion system uses a Qwen 0.6B model. If the weights aren't downloaded (they're separate from the main checkpoints), the pipeline feeds garbage data to BigVGAN. Add a guard:

# In __init__, wrap QwenEmotion loading
try:
    self.qwen_emo = QwenEmotion(
        os.path.join(self.model_dir, self.cfg.qwen_emo_path)
    )
except Exception as e:
    print(f"QwenEmotion failed: {e}")
    self.qwen_emo = None

Download the emotion model weights from the IndexTTS-2 HuggingFace repo — specifically the qwen0.6bemo4-merge/model.safetensors file (~1.1GB).

2. BigVGAN Output Normalization

Safety net for edge cases where BigVGAN outputs values slightly outside [-1, 1]:

# Before the int16 scaling line
wav_max = wav.abs().max()
if wav_max > 1.0:
    wav = wav / wav_max

Running on Apple Silicon: What Actually Works

We tested on a Mac Mini M4 with 16GB unified memory. Here's what we found:

• CPU mode works. All components (GPT, s2mel diffusion, BigVGAN) run correctly on CPU. Performance: ~80 seconds for 9 seconds of audio (RTF ~8.6x).
• MPS partially works. GPT and BigVGAN run fine on MPS. But the s2mel diffusion transformer with CFG (classifier-free guidance) at 0.7 needs ~14.5GB of MPS memory — too much for 16GB Macs.
• MPS without CFG fits but sounds worse. Setting inference_cfg_rate=0.0 halves memory usage and runs on MPS, but CFG is critical for quality.
• Full MPS with PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 — crashed the machine (swap death).

Our recommendation: Use CPU mode with the default settings (CFG=0.7, 25 diffusion steps). It's slower but produces the same quality as CUDA. For production batches, run overnight.

# Force CPU mode on Apple Silicon
import torch
torch.backends.mps.is_available = lambda: False

from indextts.infer_v2 import IndexTTS2
tts = IndexTTS2(model_dir='./checkpoints', cfg_path='./checkpoints/config.yaml')

tts.infer(
    spk_audio_prompt='voice_sample.wav',
    text='Your text here.',
    output_path='output.wav',
    verbose=True
)

Voice Sample Tips

Your reference audio matters more than any hyperparameter:

• Use 10+ seconds of clear speech (not 5s — the model needs more data to clone well)
• RMS around 0.08-0.12 is ideal. Too quiet (< 0.05) gives weak output
• Clean recordings only — no background music, no reverb, no compression artifacts
• 22050 Hz sample rate — match the model's native rate

The Debugging Lesson

Always verify the saved file, not just the model output tensor.

We spent hours looking at BigVGAN output tensors, diffusion step counts, MPS vs CPU precision, voice sample quality — everything upstream. The model was perfect. The save function was silently destroying the audio on every single run.

If your model outputs look correct in the debugger but the WAV file sounds wrong, check your save pipeline. torchaudio.save() with int16 data is the most common silent killer.