 |
ML Newsnew | past | best | rss | submit | about |
| Vq-Wav2vec: Self-Supervised Learning of Discrete Speech Representations (arxiv.org) | |
| 3 points by sel1 on Oct 16, 2019 | hide | past | pdf | discuss on HN | |
|
In plain words: It turns short slices of speech into a small set of discrete symbols by clustering the audio's learned features, so text-style pre-training tricks can run on sound. Combined with a masked-word guessing scheme, it beat systems using continuous audio features on phoneme recognition and speech transcription. Abstract · vq-wav2vec: Self-Supervised Learning of Discrete Speech RepresentationsWe propose vq-wav2vec to learn discrete representations of audio segments through a wav2vec-style self-supervised context prediction task. The algorithm uses either a gumbel softmax or online k-means clustering to quantize the dense representations. Discretization enables the direct application of algorithms from the NLP community which require discrete inputs. Experiments show that BERT pre-training achieves a new state of the art on TIMIT phoneme classification and WSJ speech recognition. | |
| add comment on HN |