3 ms·
The paper describing this work: http://arxiv.org/pdf/1507.06947v1.pdf http://arxiv.org/pdf/1507.06947v1.pdf They trained on 3 million "utterances" of average d
by nl 11y ago
The paper describing this work: http://arxiv.org/pdf/1507.06947v1.pdf http://arxiv.org/pdf/1507.06947v1.pdf
They trained on 3 million "utterances" of average duration of 4 seconds. These were distorted by noise to get 20 variations (so the training set was 60 million utterances total).
I don't understand if these were labeled somehow There's a section on clustering into 9287 phones, but it isn't clear to me if these were used as labels.
- afsina 11y agoAFAIK current ASR systems uses grapheme transcriptions for training their acoustic models. So they have the speech and the transcription "Hello World" and during training, they are automatically converted to phones kind of "hələʊ wɜ:ld" using extensive phonetic dictionaries and some algorithms. 9287 is context dependent phones. like "a" but there is a "b" on the left and "r" on the right. Theoretically for 40 phones you end up 40^3 context dependent phones. But in practice this number is much lower.