Auxiliary Loss Function for Target Speech Extraction and Recognition with Weak Supervision Based on Speaker Characteristics

Automatic speech recognition systems deteriorate in presenceof overlapped speech. A popular approach to alleviate this is targetspeech extraction. The extraction system is usually trainedwith a loss function measuring the discrepancy between the estimatedand the reference target speech. This often leads to distortionsto the target signal which is detrimental to the recognitionaccuracy. Additionally, it is necessary to have the strongsupervision provided by parallel data consisting of speech mixturesand single-speaker signals. We propose an auxiliary lossfunction for retraining the target speech extraction. It is composedof two parts: first, a speaker identity loss, forcing the estimatedspeech to have correct speaker characteristics, and second,a mixture consistency loss, making the extracted sourcessum back to the original mixture. The only supervision requiredfor the proposed loss is speaker characteristics obtainedfrom several segments spoken by the target speaker. Such weaksupervision makes the loss suitable for adapting the system directlyon real recordings. We show that the proposed loss yieldssignals more suitable for speech recognition and further, wecan gain additional improvements by adaptation to target data.Overall, we can reduce the word error rate on LibriCSS datasetfrom 27.4% to 24.0%.