pyproject.toml

[build-system]
requires = ["setuptools>=64"]
build-backend = "setuptools.build_meta"

[project]
name = "srfd"
version = "0.1.0"
description = "Speech Representation Fréchet Distance (SR-FD) loss for few-step flow-matching TTS"
readme = "README.md"
requires-python = ">=3.10"
license = { text = "Apache-2.0" }
keywords = ["text-to-speech", "flow-matching", "frechet-distance", "speech-generation", "lora"]

# Core dependency for the differentiable loss and moment/Fréchet utilities.
dependencies = [
    "torch>=2.5",
]

[project.optional-dependencies]
# Frozen content extractors (Whisper encoder + wav2vec2 CTC) and reference-stat
# computation from audio.
extractors = [
    "transformers>=4.40",
    "torchaudio>=2.5",
    "soundfile",
    "pyyaml",
]
# Seed-TTS evaluation (WER / SIM / UTMOS / DNSMOS) and significance tests.
eval = [
    "transformers>=4.40",
    "torchaudio>=2.5",
    "soundfile",
    "scipy",
    "numpy",
    "jiwer",
    "librosa",
]
test = [
    "pytest",
    "pyyaml",
]

[project.urls]
Paper = "https://arxiv.org/abs/2607.06027"
Source = "https://github.com/voidful/srfd-tts"
Documentation = "https://github.com/voidful/srfd-tts#readme"

[tool.setuptools]
packages = ["srfd"]

[tool.pytest.ini_options]
testpaths = ["tests"]