InterSpeech 2021

Whispering is the natural choice of communication when one wants to interact quietly and privately. Due to vast differences in acoustic characteristics of whisper and natural speech, there is drastic degradation in the performance of whisper speech when decoded by the Automatic Speech Recognition (ASR) system trained on neutral speech. Recently, to handle this mismatched train and test scenario Denoising Autoencoders (DA) are used which gives some improvement. To improve over DA performance we propose another method to map speech from whisper domain to neutral speech domain via Joint Variational Auto-Encoder (JVAE). The proposed method requires time-aligned parallel data which is not available, so we developed an algorithm to convert parallel data to time-aligned parallel data. JVAE jointly learns the characteristics of whisper and neutral speech in a common latent space which significantly improves whisper recognition accuracy and outperforms traditional autoencoder based techniques. We benchmarked our method against two baselines, first being ASR trained on neutral speech and tested on whisper dataset and second being whisper test set mapped using DA and tested on same neutral ASR. We achieved an absolute improvement of 22.31% in Word Error Rate (WER) over the first baseline and an absolute 5.52% improvement over DA.