Aligning an existing transcript

[samuelbradshaw]

Does this library provide a way to align an existing transcript to an audio recording?

[LauraGPT]

Yes. For an existing transcript, use FunASR's fa-zh forced-alignment model:

from funasr import AutoModel

model = AutoModel(model="fa-zh")
result = model.generate(
    input=("audio.wav", "transcript.txt"),
    data_type=("sound", "text"),
)
print(result)

The second input is the transcript to align, and the result contains the predicted token/character timing. The transcript should correspond closely to the spoken audio; large insertions, omissions, or a different language/model domain can reduce alignment quality.

fa-zh maps to iic/speech_timestamp_prediction-v1-16k-offline. It is the dedicated Chinese timestamp/forced-alignment path, not a generic multilingual aligner. Fun-ASR-Nano's own CTC timestamp stage aligns its decoded output and is a different workflow.

Official example: Timestamp Prediction.