How to Install IndexTTS2 on Mac (Apple Silicon M1/M2/M3/M4 Guide)

IndexTTS2 is one of the best open-source text-to-speech models for voice cloning — but the official setup assumes you have an NVIDIA GPU with CUDA. If you're on a Mac with Apple Silicon, you'll hit a few walls.

This guide walks you through installing IndexTTS2 on Mac (M1, M2, M3, M4), getting it running in CPU mode, and fixing the critical bugs that make the output sound like static.

1. Prerequisites

Before starting, you'll need:

• Mac with Apple Silicon (M1, M2, M3, or M4)
• 16 GB RAM minimum (32 GB recommended for MPS acceleration)
• 15 GB free disk space for model checkpoints + environment
• Homebrew installed (/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)")
• Git and Git LFS

Install Python 3.11

IndexTTS2 works best with Python 3.11. Python 3.12+ may cause issues with numba and other dependencies.

brew install [email protected] git-lfs
git lfs install

2. Installation

cd ~/projects  # or wherever you keep projects
git clone https://github.com/IndexTeam/IndexTTS-2.git indexTTS2
cd indexTTS2

python3.11 -m venv ~/indextts2-venv
source ~/indextts2-venv/bin/activate

pip install --upgrade pip
pip install -e .

This will install PyTorch (with MPS support), transformers, BigVGAN, and all other dependencies. On Apple Silicon, PyTorch will automatically include MPS backend support.

3. Download Model Checkpoints

IndexTTS2 needs several model files. The main checkpoints should download automatically on first run, but the Qwen emotion model needs manual download.

This is the most commonly missed step. Without the emotion model, the pipeline feeds garbage data to the vocoder.

# Download from HuggingFace
pip install huggingface_hub

python3 -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
    repo_id='IndexTeam/IndexTTS-2',
    filename='qwen0.6bemo4-merge/model.safetensors',
    local_dir='./checkpoints'
)
print('Qwen emotion model downloaded!')
"

Verify the file exists:

ls -lh checkpoints/qwen0.6bemo4-merge/model.safetensors
# Should be ~1.1 GB

4. Critical Fix: The torchaudio.save Bug

This is not a BigVGAN issue or an MPS issue — the model generates perfect audio, but the save function corrupts it. Read our full analysis.

# FIND this line:
torchaudio.save(output_path, wav.type(torch.int16), sampling_rate)

# REPLACE with:
torchaudio.save(output_path, (wav / 32767.0).clamp(-1.0, 1.0).float(), sampling_rate, bits_per_sample=16)

This saves properly normalized float32 audio instead of broken int16.

5. Additional Fixes

Fix 1: Qwen Emotion Model Guard

If the emotion model fails to load for any reason, the whole pipeline crashes. Add a try/except guard.

In indextts/infer_v2.py, find:

self.qwen_emo = QwenEmotion(os.path.join(self.model_dir, self.cfg.qwen_emo_path))

Replace with:

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

Fix 2: BigVGAN Output Normalization

Safety net for rare cases where BigVGAN outputs values slightly outside [-1, 1]. Add before the torch.clamp(32767 * wav line:

# Add BEFORE the torch.clamp line:
wav_max = wav.abs().max()
if wav_max > 1.0:
    wav = wav / wav_max

6. Preparing a Voice Sample

Your voice reference is the most important input. IndexTTS2 clones the voice characteristics from this sample.

• Duration: 10 seconds minimum (longer is better)
• Format: WAV, 22050 Hz, 16-bit, mono
• Content: Clear speech, no background noise or music
• Volume: RMS around 0.08-0.12 (not too quiet, not clipping)

Convert any audio to the right format with ffmpeg:

brew install ffmpeg

# Convert and extract a 10-second clip
ffmpeg -i input.mp3 -ss 30 -t 10 -ar 22050 -ac 1 -acodec pcm_s16le voice_sample.wav

Check the sample quality:

python3 -c "
import librosa, numpy as np
y, sr = librosa.load('voice_sample.wav', sr=None)
rms = np.sqrt(np.mean(y**2))
print(f'SR: {sr}, Duration: {len(y)/sr:.1f}s, RMS: {rms:.4f}, Peak: {np.abs(y).max():.4f}')
if rms < 0.05: print('⚠️  Too quiet — consider amplifying')
if rms > 0.3: print('⚠️  Very loud — may clip')
if sr != 22050: print('⚠️  Sample rate should be 22050 Hz')
"

7. Running Inference

source ~/indextts2-venv/bin/activate
cd ~/projects/indexTTS2

python3 -c "
import torch
# Force CPU mode on Apple Silicon (recommended for 16GB Macs)
torch.backends.mps.is_available = lambda: False

from indextts.infer_v2 import IndexTTS2

# Load model
tts = IndexTTS2(model_dir='./checkpoints', cfg_path='./checkpoints/config.yaml')

# Generate speech
tts.infer(
    spk_audio_prompt='voice_sample.wav',
    text='Hello, this is a test of IndexTTS2 running on Apple Silicon.',
    output_path='output.wav',
    verbose=True
)
print('Done! Check output.wav')
"

First run will be slower as PyTorch compiles operations. Subsequent runs are faster.

8. MPS vs CPU: What Works

Apple's Metal Performance Shaders (MPS) can accelerate some components but not all. Here's what we tested on a Mac Mini M4 (16GB):

9. Performance Benchmarks

Tested on Mac Mini M4 (16GB), CPU mode, default settings (CFG=0.7, 25 diffusion steps):

Not real-time, but perfectly usable for batch generation. A 5-minute audio script takes about 45 minutes to generate.

10. Troubleshooting

Output sounds like static / heavily distorted

Apply the torchaudio.save fix. This is the #1 cause of bad audio on non-CUDA setups.

OOM crash or system freeze on Apple Silicon

Force CPU mode: torch.backends.mps.is_available = lambda: False

"QwenEmotion" error on startup

Download the emotion model weights — see Step 4.

Very slow first run

Normal. PyTorch compiles operations on first execution. BigVGAN also downloads its weights (~100MB) from NVIDIA's HuggingFace repo on first use.

Disk full errors with MPS

MPS caches compiled graph files to disk. Ensure at least 5GB free space. Clear the cache: rm -rf ~/Library/Caches/com.apple.metal/

Output voice doesn't match reference

Check your voice sample: 10+ seconds, 22050 Hz, clear speech, no background noise. Shorter or noisy samples produce poor clones.

"numba" or "llvmlite" installation errors

Use Python 3.11 specifically: python3.11 -m venv ~/indextts2-venv