Transfer Learning for Tandem ASR Feature Extraction
Tandem automatic speech recognition (ASR), in which one or an ensemble of multi-layer perceptrons (MLPs) is used to provide a non-linear transform of the acoustic parameters, has become a standard technique in a number of state-of-the-art systems. In this paper, we examine the question of how to transfer learning from out-of-domain data to new tasks.
Our primary focus is to develop tandem features for recognition of speech from the meetings domain. We show that adapting MLPs originally trained on conversational telephone speech leads to lower word error rates than training MLPs solely on the target data. Multi-task learning, in which a single MLP is trained to perform a secondary task (in this case a speech enhancement mapping from farfield to nearfield signals) is also shown to be advantageous.
We also present recognition experiments on broadcast news data which suggest that structure learned from English speech can be adapted to Mandarin Chinese. The performance of tandem MLPs trained on 440 hours of Mandarin speech with a random initialization was achieved by adapted MLPs using about 97 hours of data in the target language.
Unable to display preview. Download preview PDF.
- 1.Hermansky, H., Ellis, D., Sharma, S.: Tandem connectionist feature stream extraction for conventional hmm systems. In: Proc ICASSP, Istanbul, Turkey, vol. III, pp. 1635–1638 (2000)Google Scholar
- 3.Stolcke, A., Grezl, F., Hwang, M.Y., Lei, X., Morgan, N., Vergyri, D.: Cross-domain and cross-language portability of acoustic features estimated by multilayer perceptrons. In: Proc. ICASSP, Toulouse, France (2006)Google Scholar
- 4.Zheng, J., Çetin, O., Hwang, M.Y., Lei, X., Stolcke, A., Morgan, N.: Combining discriminative feature, transform, and model training for large vocabulary speech recognition. In: Proc. ICASSP, Honolulu (2007)Google Scholar
- 6.Hain, T., Burget, L., Dines, J., Garau, G., Karafiat, M., Lincoln, M., Vepa, J., Wan, V.: The AMI meeting transcription system: Progress and performance. In: NIST RT 2006 Workshop (2006)Google Scholar
- 7.Hermansky, H.: TRAP-TANDEM: Data-driven extraction of temporal features from speech. In: IDIAP-RR 50, IDIAP, Martigny, Switzerland (2003)Google Scholar
- 9.Chen, B., Zhu, Q., Morgan, N.: Learning long term temporal feature in LVCSR using neural networks. In: Proc. ICSLP, pp. 612–615 (2004)Google Scholar
- 10.Zhu, Q., Stolcke, A., Chen, B.Y., Morgan, N.: Using MLP features in SRI’s conversational speech recognition system. In: Proc. Eurospeech, Portugal (2005)Google Scholar
- 11.Janin, A., Stolcke, A., Anguera, X., Boakye, K., Çetin, O., Frankel, J., Zheng, J.: The ICSI-SRI spring 2006 meeting recognition system. In: Proc. MLMI, Washington DC, USA (2006)Google Scholar
- 12.Hwang, M.Y., Wang, W., Lei, X., Zheng, J., Çetin, O., Peng, G.: Advances in Mandarin broadcast speech recognition. In: Proc.Interspeech, Antwerp (2007)Google Scholar