It’s no secret that the science of speech recognition has come a long way since IBM introduced its first speech recognition machine in 1962. As the technology has evolved, speech recognition has become increasingly embedded in our everyday lives with voice-driven applications like Amazon’s Alexa, Apple’s Siri, Microsoft’s Cortana, or the many voice-responsive features of Google. From our phones, computers, watches and even our refrigerators, each new voice-interactive device that we bring into our lives deepens our dependence on artificial intelligence (AI) and machine learning.
source:
The Role of Artificial Intelligence and Machine Learning in Speech Recognition | Rev
Eduard Babulak
Speech recognition and audio processing are two areas in which AI has made significant advancements in recent years. These advances are enabling the development of more accurate and efficient simultaneous translation systems. Here are a few applications of AI in this area:
These are just a few examples of how AI is being used to improve simultaneous translation systems. As AI technology continues to evolve, it is likely that these systems will become even more accurate and efficient, enabling people to communicate across languages more seamlessly and effectively.
In very simplistic terms a speech recognition engine is usually divided into three components:
1) The acoustic model
2) The pronunciation dictionary
3) The language model
The traditional speech recognition system consisted of handing to the acoustic model a processed singal in the form of MFCCs which where then used to train an GMM-HMM-HMM model (the first HMM modeled subword units while the second was trained to consider multi word units.
The traditional language model has since been replaced by NN-HMM-HMM models where the NN can be a TDNN[1] (time delay neural network).
Lately, more advanced architectures have gone as far as removing the MFCC acoustic features in favor of W2V[2] and relying mostly on Transformer NNs that can handle multiple languages[3]. These newer models are trained with minimal supervision.
Once the audio is transcribed then the transcription is passed to a digital assistance that processes the command. Further, this is just a very incomplete summary just to highlight some of the research.
References
[1]Peddinti, V., Povey, D., & Khudanpur, S. (2015). A time delay neural network architecture for efficient modeling of long temporal contexts. In Sixteenth annual conference of the international speech communication association.
[2] Baevski, A., Zhou, Y., Mohamed, A., & Auli, M. (2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33, 12449-12460.
[3]Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2022). Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356.
- Badges
- Science topic