| [project] | |
| name = "prosody_explorer" | |
| version = "0.1.0" | |
| description = "Prosody features to mel to text -- the Prosody2text pipeline" | |
| requires-python = ">=3.10" | |
| dependencies = [ | |
| # The feature extractor. amfm-decompy is pinned exactly: f0_interp builds | |
| # a pYAAPT.PitchObj without __init__ and reads class constants off it, so | |
| # a release that moves those to instance attributes breaks gap | |
| # interpolation at runtime rather than at import. | |
| "amfm-decompy==1.0.12.2", | |
| # Praat's intensity analysis. Wheels exist for Linux, macOS and Windows. | |
| "praat-parselmouth>=0.4.7", | |
| "numpy>=1.26", | |
| "scipy>=1.11", | |
| # The conformer is the trained torch module, loaded from safetensors. | |
| "torch>=2.8", | |
| "safetensors>=0.4", | |
| # transformers is pinned to the 5.x line the reference was measured | |
| # against: the beam capture monkeypatches GenerationMixin.generate, so a | |
| # major version is a real compatibility boundary. | |
| "transformers>=5.20,<6", | |
| "tokenizers>=0.23,<0.24", | |
| "huggingface-hub>=0.25", | |
| # Reading an audio file by path. Ships its own libsndfile, so unlike | |
| # torchcodec it needs no system codec. | |
| "soundfile>=0.12", | |
| ] | |
| [project.optional-dependencies] | |
| # Only for evaluate_dataset.py. Kept out of the base install because | |
| # `datasets` pulls a large tree the pipeline itself never touches. | |
| # | |
| # soundfile rather than torchcodec, deliberately: datasets 5.x decodes audio | |
| # through torchcodec, which dlopens system FFmpeg libraries and so fails | |
| # with "libavutil.so not found" on a machine that has only wheels. The | |
| # script asks datasets for undecoded bytes and reads them with soundfile, | |
| # which ships its own libsndfile. | |
| eval = [ | |
| "datasets>=3.0", | |
| "jiwer>=4.0.0", | |
| "soundfile>=0.12", | |
| ] | |
| [build-system] | |
| requires = ["hatchling"] | |
| build-backend = "hatchling.build" | |
| # The modeling files sit at the repo root, which is the Hub's own convention | |
| # for a custom architecture. Mapping them into a package here lets the same | |
| # files be imported normally by anything that installs this -- so the Space | |
| # can `from prosody_explorer.features import ...` at module level rather | |
| # than reaching into the revision-hashed path trust_remote_code creates. | |
| [tool.hatch.build.targets.wheel] | |
| only-include = [ | |
| "features.py", "f0_interp.py", "_yaapt_patch.py", | |
| "speech.py", "conformer.py", "conformer_model.py", | |
| "whisper_decoder.py", "configuration_prosody.py", "modeling_prosody.py", | |
| ] | |
| sources = { "" = "prosody_explorer" } | |