speechrecognition

Speech recognition is the process of converting spoken language into written text. Modern systems are designed to handle variations in speaker voice, accent, speaking rate, and background noise, and to operate in real time or near real time.

Historically, speech recognition progressed from template matching and dynamic time warping to statistical modeling. Hidden Markov models with Gaussian mixture models dominated the era from the 1980s through the 2010s. The rise of deep learning brought neural network–based approaches, including hybrid HMM–DNN systems and, more recently, end-to-end architectures such as connectionist temporal classification and attention-based sequence-to-sequence models. These models can learn representations directly from audio, often using spectrogram or waveform inputs.

A typical automatic speech recognition system comprises an acoustic model, a pronunciation lexicon or grapheme-to-phoneme component, a language model, and a decoder. The acoustic model maps audio features to phonetic states, while the language model assigns probabilities to word sequences. The decoder searches for the most likely transcription given the acoustic evidence. Features commonly used include MFCCs or spectrograms; many modern systems employ end-to-end neural networks that operate on raw or minimally processed audio.

Performance is commonly evaluated using word error rate, which accounts for substitutions, insertions, and deletions relative to a reference transcript. Large public datasets such as LibriSpeech, Switchboard, and TED-LIUM are used for benchmarking. Privacy, bias, and representation of diverse languages and dialects are important considerations in both research and deployment.

Applications of speech recognition include voice assistants, transcription services, real-time captioning, and accessibility tools. It enables hands-free control and content indexing but continues to face challenges related to noise, accents, multilingual content, and resource constraints for low-resource languages.