# Core dependencies for tiny-audio model inference
# This file is pushed to HuggingFace for model repository

# Transcription and word timestamps (Qwen3-ForcedAligner) need transformers >= 5.17.
# Speaker diarization (return_speakers=True) uses Nemotron-3-Diarization, which is in
# transformers main after 5.17 -- pinned to the commit the pipeline was verified on.
transformers @ git+https://github.com/huggingface/transformers@91ea1192a00e0ebb684400434375fbd2345c8449
torch
torchaudio
peft
# Nemotron's streaming feature extractor (return_speakers=True) imports it.
librosa