# ---------------------------------------------------------------------------
# Install in two steps with pip (or one with uv — see README):
#
#   pip install -r requirements.txt
#   pip install --no-deps "chatterbox-tts @ git+https://github.com/resemble-ai/chatterbox@ed27b95"
#
# Why --no-deps: chatterbox's own metadata pins numpy<1.26, which has no wheel
# for Python 3.12 — a plain `pip install chatterbox-tts` tries to build numpy
# 1.25 from source and dies on the removed `distutils`. The code itself runs
# fine on NumPy 2.x; the real ceiling is numba's, below. So chatterbox's
# runtime dependencies are listed here explicitly, from its pyproject at that
# commit, minus `gradio` (a demo UI it never imports on the Turbo path, and a
# quarter of the install by itself).
# ---------------------------------------------------------------------------

# Windows + NVIDIA: PyPI's Windows torch is CPU-only. Run this *first*, then
# the two steps above; pip keeps the CUDA build since the version matches:
#
#   pip install torch==2.6.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
#
# Linux and macOS need nothing extra. (`uv run` handles this automatically.)

# --- mynah ---
fastapi>=0.115
hf_xet
uvicorn[standard]>=0.32
python-multipart>=0.0.9
soundfile>=0.12
imageio-ffmpeg>=0.5

# resemble-perth (the watermarker) imports `pkg_resources`, which setuptools
# removed in 81 — and uv/venv environments do not ship setuptools at all. It
# swallows the ImportError and leaves the watermarker as None, so the failure
# surfaces later as "'NoneType' object is not callable" at model load.
setuptools<81

# --- chatterbox-tts @ ed27b95, runtime dependencies ---
# numba (via librosa) refuses NumPy 2.4+; 1.26 is the first with 3.12 wheels.
numpy>=1.26,<2.4
librosa==0.11.0
s3tokenizer
torch==2.6.0
torchaudio==2.6.0
transformers==4.46.3
diffusers==0.29.0
resemble-perth==1.0.1
conformer==0.3.2
safetensors==0.5.3
spacy-pkuseg
pykakasi==2.3.0
pyloudnorm
omegaconf
huggingface-hub