Forced Alignment

[qo4on]

Hi!
I have an audio and a text of it. Can NeMo return timestamps for every word in the text file? I believe AlignerModel is what I need. What is the right way to use it?

import nemo.collections.tts as nemo_tts
tts_models = [model for model in dir(nemo_tts.models) if model.endswith("Model")]

['AlignerModel',
 'DegliModel',
 'EDMel2SpecModel',
 'FastPitchHifiGanE2EModel',
 'FastPitchModel',
 'FastSpeech2HifiGanE2EModel',
 'FastSpeech2Model',
 'GlowTTSModel',
 'GriffinLimModel',
 'HifiGanModel',
 'MelGanModel',
 'MelPsuedoInverseModel',
 'SqueezeWaveModel',
 'Tacotron2Model',
 'TalkNetDursModel',
 'TalkNetPitchModel',
 'TalkNetSpectModel',
 'TwoStagesModel',
 'UniGlowModel',
 'WaveGlowModel']

[qo4on]

Is there anyone alive?

[PatrykNeubauer]

If I'm not mistaken, the aligner model is designed for use with TTS models as an alignment learning framework.
Not sure, if there's a simple way to get what you want out of NeMo, but Montreal Forced Aligner can do exactly what you want easily.
It's even used by NeMo in it's scripts for ljspeech preprocessing

[titu1994]

Yup, we're alive. MFA is very well supported for exactly this task. We're working on aligners, but that will take time.