This checkpoint is a wav2vec2-large model that is useful for generating transcriptions with punctuation. It is intended for use in building transcriptions for TTS models, where punctuation is very important for prosody.
The model gets a respectable WER of 4.45% on the librispeech validation set. The baseline,
facebook/wav2vec2-large-robust-ft-libri-960h, got 4.3%.
Since the model was fine-tuned on clean audio, it is not well-suited for noisy audio like CommonVoice (though I may upload a checkpoint for that soon too). It still does pretty good, though.
The vocabulary is uploaded to the model hub as well
- Downloads last month