Connectionist Temporal Classification (CTC)
The loss function that let a neural network transcribe speech without being told where each sound begins.
Why It Was Important
Training a network on speech previously required knowing which audio frame corresponded to which phoneme — an alignment that had to be produced by a hidden Markov model first. CTC removed that requirement by summing over every possible alignment, so a single network could map audio straight to text. It is the technical hinge on which speech recognition swung away from the HMM pipeline that had run it for thirty years and toward the end-to-end neural systems behind Deep Speech, Siri and Whisper.
Who Invented It
Alex Graves, Santiago Fernández, Faustino Gomez, Jürgen Schmidhuber
Developed at IDSIA, in the group that had produced the LSTM.
Applications
- Speech Recognition
- Handwriting Recognition
- Sequence Labelling
- Lip Reading
Key Papers
- Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks
Alex Graves et al. · ICML 2006