The Deep Learning Revolution in Automatic Speech Recognition
summary
Describes how speech systems move from GMM-HMM acoustic models to shared-parameter deep neural networks trained discriminatively, enabled by GPUs, producing a leap from about 23.6% to 16.1% WER (~32% relative improvement). Replacing DNNs with LSTM RNNs gives another ~15% relative gain down to ~9.9% on Switchboard, and ensemble systems of many models reach ~5.1% WER, plus sequence-to-sequence models that collapse acoustic model, lexicon and language model into one net.
AI-written summary · about
chapters (10)
- 0:00Introduction to the topic
- 3:24Speech recognition basics
- 8:00System architecture
- 14:56GMM and HMM foundations
- 20:53The Switchboard task
- 24:27The neural network era
- 30:14RNNs and LSTMs
- 34:01Neural language models
- 39:10Future directions
- 42:32Q&A session
related talks
27mThe Story of End to End Models in Deep Learning
25mFast deep recurrent net training
25mBeyond Transcription: Building Voice AI That Understands Conversations
21mDeveloping Machine Learning Based Speech Enhancement Models for Teams and Skype
42mdata.bythebay.io: Shubho Sengupta,
56mRandom Embeddings, Matrix-valued Kernels and Deep Learning
06mDeepRange: Acoustic Ranging via Deep Learning
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python