Scenario 1

Need to perform ASR over short chunks and store in nemo_manifest

import os

from vac_aligner.dto import ASRConfig
from vac_aligner.asr import ASR_MAPPING, ASR

language = "hy"  #  Armenain
model_name = ASR_MAPPING[language] # "Yeroyan/stt_arm_conformer_ctc_large"
asr_config = ASRConfig(hf_token=os.environ['HF_TOKEN'], batch_size=24)
asr = ASR(model_name, asr_config)
asr.run(
    save_dir="where/to/save/asr/predictions", # .txt(s)
    save_manifest="path/to/save/manifest.json",
    wav_files="path/to/wav/files",
    test_manifest='path/to/manifest/with/predictions.json'
)

Scenario 2

You already have predictions manifest and/or long transcript, and want to run the Matching to obtain the correct chunk texts to replace ASR predictions.

from vac_aligner.matching import ArmenianAlignerVAC

output_file = 'path/to/save/combined/transcript.txt'
predictions_manifest = 'path/to/manifest/with/predictions.json'
chunks, combined_transcript = ArmenianAlignerVAC.combine_transcript(predictions_manifest,
                                                                    output_file,
                                                                    ending_punctuations="․,։")

and then matching

from vac_aligner.matching import ArmenianAlignerVAC

target_base = 'path/to/save/artifacts'
matches_sorted = ArmenianAlignerVAC(combined_transcript,
                                    chunks, output_file.replace(".txt", ".json"),
                                    target_base=target_base).align(0.35)

and get the benchmark

from vac_aligner.matching.benchmark_on_mcv import Benchmark

benchmark = Benchmark(target_base, predictions_manifest)
stats = benchmark.get_benchmark()
benchmark.analyze_and_save_benchmark(stats, output_file)