montague_archiving

ASR scripts for the Montague biography project

Main file is annotate.py

usage: python annotate.py [-h] [-i INPUT] [-s {asr-only,asr-first,drz-first,multitier}] [-n NUM_SPEAKERS] [-m ASR_MODEL] [-d DRZ_MODEL] [-D DEVICE] [-c CHUNK_LENGTH_S] [-w] [-r] [-b ASR_BATCH_SIZE]

optional arguments:

• -h, --help show this help message and exit
• -i INPUT, --input INPUT .wav file or directory of .wav files to annotate
• -s {asr-only,asr-first,drz-first,multitier}, --strategy {asr-only,asr-first,drz-first,multitier} Specify what pipeline to use for annotation. asr-first (default) will run Whisper first then diarization with PyAnnote, and the PyAnnote diarization will be used to decide the speaker identityfor each chunk output by Whisper. asr-only will run Whisper without performing speaker diarization. drz-first will run diarization with PyAnnote first and then send each speaker turn to Whisper for transcription. This will likely result in poor turn boundaries, so asr-first should generally be preferred. multitier will output an .eaf file with Whisper annotations and PyAnnote speaker turns on separate tiers, and is mostly useful for debugging and understanding the decisions that ASR and diarization are making independently.
• -n NUM_SPEAKERS, --num_speakers NUM_SPEAKERS Number of speakers in file. Default 2.
• -m ASR_MODEL, --asr_model ASR_MODEL ASR model path. Default is Nvidia Parakeet.
• -d DRZ_MODEL, --drz_model DRZ_MODEL DRZ model path. Default is pyannote/speaker-diarization-3.1.
• -D DEVICE, --device DEVICE Device to run GPU on. Defaults to cuda:0 if available else cpu.
• -c CHUNK_LENGTH_S, --chunk_length_s CHUNK_LENGTH_S Chunk size to input to Whisper pipeline (default 30s). Note that this will not affect the size of chunks OUTPUT by Whisper, so this should only be modified if there is not enough memory to process 30s at a time.
• -w, --return_word_timestamps Whisper by default chunks speech more or less into utterances. Use this option to chunk by word, which may give more precise time accuracy When detecting speaker changes. (I recommend using this option.)
• -r, --recursive If input is a directory, will search for files .wav files recursively
• -b ASR_BATCH_SIZE, --asr_batch_size ASR_BATCH_SIZE Inference batch size for ASR. Default 8.

For each input .wav file, saves a .eaf and .txt file with annotations for that recording with the same path as the input file.