Speech recognition

Edit · View history

Speech recognition

Speech recognition is the technology that enables a computer or device to identify and process human speech into text or commands. It is also known as automatic speech recognition (ASR) or speech-to-text. The field combines linguistics, computer science, and signal processing, and has become a common interface for virtual assistants, transcription services, and accessibility tools.

Early systems in the 1950s could recognize only single digits spoken by a single voice. Modern ASR systems use deep neural networks trained on vast corpora of recorded speech. They are capable of handling multiple languages, accents, and noisy environments. Common applications include dictation software, voice-controlled assistants like Siri and Alexa, call‐centre automation, and real‑time captioning.

History

The first speech recognizer, Audrey, was built by Bell Laboratories in 1952 and could recognize digits 0‑9. In the 1960s, IBM introduced Shoebox which understood 16 English words. The 1970s saw the development of hidden Markov models (HMMs), which became the dominant approach for decades. The Darpa speech‐understanding project in the 1970s funded research leading to systems that could handle continuous speech with limited vocabularies.

In the 1990s, commercial products like Dragon NaturallySpeaking brought dictation to personal computers. The 2010s witnessed a shift to deep learning, with companies such as Google, Microsoft, and Amazon deploying large‑scale neural network models. The use of recurrent neural networks, later transformers and attention mechanisms, dramatically reduced word error rates.

Features

Modern speech recognition systems typically include:

• Language modeling to predict the most likely word sequences.
• Acoustic modeling to map audio signals to phonetic units.
• Endpoint detection to identify when speech starts and stops.
• Adaptation to a user’s voice and accent over time.
• Noise suppression and echo cancellation for use in varied environments.

Challenges

Key challenges remain: handling diverse dialects, overlapping speakers, and low‑resource languages. Privacy concerns also arise, as many systems process audio in the cloud. Researchers continue to explore end‑to‑end models and self‑supervised learning to improve robustness.