Title:AN ATTENTION-BASED NEURAL NETWORK APPROACH FOR SINGLE CHANNEL SPEECH ENHANCEMENT
-
What’s main claim? Key idea?
In this paper, the authors explore the attention-based neural network structures for improving the performance of speech enhancement. This paper adopt attention mechanism on LSTM-RNN models. Specifically, taking noisy spectrum as input, this model is composed of an LSTM based encoder, an attention mechanism and a speech generator, resulting in enhanced spectrum.
-
Is there code available? Data?
No code
Data: randomly select a clean speech file from a multiple-speaker speech corpus which has 21407 utterances (about 24.5 hours) and a noise file from the Musan corpus.
-
Is the idea neat? Is it counter-intuitive?
It is intuitive to use attention mechanism in speech enhancement as humans are able to focus on the important speech components in an audio stream with “high attention” while perceiving the unimportant region (e.g., noise or interference) in “low attention”, and thus adjust the focal point over time.
-
Is the experimentation good? Manual tuning?
Compared with OM-LSA and an LSTM approach without attention mechanism.
Loss function: mean square error (MSE).
The learning rate is set to 0.0005 at the beginning.
Evaluation criteria: PESQ and STOI.
-
Is it useful to my work e.g. product dev?
Yes, recently, I was trying to apply attention to speech enhancement. However, the effect of using attention directly in speech enhancement is not good, so I am learning some better application methods.