4 ms·
I'm curious how much of the accuracy gains come from only having to run the decoder when the LSTM emits a phoneme rather than for each 10 ms frame, which presum
by gok 11y ago
I'm curious how much of the accuracy gains come from only having to run the decoder when the LSTM emits a phoneme rather than for each 10 ms frame, which presumably allows the language model search to be much more aggressive.
- nl 11y agoWhere are you seeing that? In the paper[1] it says: Acoustic features are generated every 10ms, but are concatenated and downsampled for input to the network: 8 frames are stacked for unidirectional (top) and 3 for bidirectional models (bottom). [1] http://arxiv.org/pdf/1507.06947v1.pdf http://arxiv.org/pdf/1507.06947v1.pdf
- gok 11y agoI'm not talking about the input, but the output: "…predicted word sequence where the word with highest prob- ability is taken ignoring repetitions and the blank label with no language model or decoding." Which I took to mean: when the acoustic model emits a blank symbol, they don't run the decoder again until a non-blank symbol comes out.
- afsina 11y agoI think that is just a technology showcase that this technique is so good that for some small-mid size pronunciation dictionary, it can work quite well without using a language model. Similar works were reported recently (ASR systems without language models or with character models). But seems like real system is still using a graph generated with a language model and context dependent phones.