Speech Recognition is the technology of converting spoken language, represented as an acoustic signal, into a machine-readable representation, most commonly a sequence of written words or tokens. It is often called Automatic Speech Recognition, abbreviated ASR, and is closely related to speech-to-text systems, although the terms can differ in scope depending on the application. A speech-recognition system receives audio captured by a microphone or another recording device, processes the acoustic waveform and estimates what linguistic content was spoken.
The task is challenging because speech varies substantially between speakers, accents, dialects, speaking rates, recording environments, microphones, background noises and communication channels. The system must distinguish speech from non-speech sounds, identify acoustic patterns associated with linguistic units, infer probable words and resolve ambiguities using contextual information. Contemporary systems generally use machine-learning models trained on large speech datasets, but speech recognition has developed through several generations of signal-processing, statistical and neural techniques.
Early speech-recognition research relied heavily on signal-processing methods and explicit models of phonetics and linguistics. Systems typically extracted acoustic features from short sections of audio and used statistical models to estimate phonetic or word sequences. Hidden Markov Models became particularly influential because they provided a probabilistic framework for representing temporal sequences, while Gaussian mixture models were widely used to model acoustic observations.
A separate language model estimated the probability of word sequences, allowing the system to prefer linguistically plausible interpretations when acoustic evidence was ambiguous. Systems were often trained for specific speakers, vocabularies or domains, and performance could degrade significantly when real-world conditions differed from training conditions. The later adoption of deep neural networks improved acoustic modelling, followed by sequence-to-sequence architectures, connectionist temporal classification, attention mechanisms, recurrent networks and transformer-based architectures. These developments increasingly enabled systems to learn representations directly from large quantities of speech data.
The basic speech-recognition problem can be represented as estimating the most probable linguistic sequence given an observed acoustic signal. The audio waveform is commonly divided into short overlapping frames, from which models derive representations that capture relevant spectral and temporal characteristics. Traditional systems used features such as Mel-frequency cepstral coefficients, while modern neural systems may learn useful representations directly from raw or minimally processed audio.
Acoustic information alone is often insufficient because multiple words can sound similar, and conversational speech may contain reductions, disfluencies and incomplete phrases. A language model or contextual component therefore contributes information about which word sequences are plausible. Modern end-to-end systems may integrate acoustic and linguistic processing into a unified neural architecture, reducing the need for separately engineered components while still relying on extensive training data and computational resources.
Speech recognition differs from speaker recognition and speech synthesis. Speaker recognition determines or verifies who is speaking, whereas speech recognition attempts to determine what was said. Speech synthesis performs the reverse general transformation by generating an acoustic signal from linguistic information. Speech understanding goes further than transcription by attempting to infer the meaning, intent or requested action contained in the spoken input.
Voice assistants often combine all of these technologies with additional components: speech recognition converts audio into text or linguistic representations, natural-language understanding interprets the request, a decision or retrieval component determines the response, and speech synthesis may vocalise the result. Automatic speech recognition can therefore function as one stage within a much larger spoken-dialogue system rather than constituting the complete voice interface.
Modern speech-recognition systems are designed to handle conditions that earlier systems struggled with, including multiple accents, spontaneous conversation, overlapping speakers and background noise. Techniques such as data augmentation, noise simulation, beamforming, source separation and robust acoustic modelling can improve performance in difficult environments. Speaker adaptation can help models account for individual vocal characteristics, while multilingual and code-switching models attempt to recognise multiple languages or transitions between languages.
Diarisation can divide a recording into speaker-specific segments, allowing a transcription to indicate which participant spoke at different times. Punctuation and capitalisation models can transform raw recognised tokens into more readable text, while timestamping associates words or segments with positions in the audio. These features are particularly useful for meetings, interviews, call-centre recordings, lectures, media archives and accessibility applications.
Speech recognition is widely used in dictation, transcription, accessibility tools, customer-service systems, automotive interfaces, meeting software, voice-controlled devices, search systems and media processing. In professional transcription, a recognised transcript may be corrected by a human editor because even highly capable systems can make errors involving names, technical vocabulary, numbers, accents or ambiguous phrases. Domain adaptation is therefore important: a medical transcription system, for example, requires appropriate terminology and training data distinct from that required for everyday conversational speech.
Recognition accuracy is commonly measured using Word Error Rate, which compares substitutions, deletions and insertions against a reference transcription. Character Error Rate can be useful for languages or applications where word segmentation behaves differently. Aggregate accuracy figures can conceal substantial variation across speakers, languages, demographic groups, acoustic environments and vocabulary types, making subgroup and condition-specific evaluation important.
Speech-recognition errors can arise from acoustic ambiguity, poor recording quality, background noise, reverberation, unusual accents, rapid speech, code-switching, domain-specific terminology, named entities and insufficient training data. Homophones are especially challenging because acoustically similar sequences may correspond to different written words, requiring contextual information to determine the intended transcription. Proper names and rare terms can have high error rates because they occur infrequently in training data.
Privacy is another important consideration because voice recordings may contain personal, confidential or commercially sensitive information. Systems that process speech remotely can introduce additional data-governance considerations concerning storage, retention, access and transmission. Real-world deployments therefore need to consider not only recognition accuracy but also latency, computational requirements, privacy, security, accessibility, supported languages and the consequences of transcription errors.
The development of speech recognition has moved from highly constrained systems towards general-purpose models capable of processing diverse speech. Self-supervised learning has become important because large quantities of unlabelled audio can be used to learn useful acoustic representations before models are adapted to labelled transcription tasks. Large multilingual models can share representations across languages and sometimes support languages for which relatively little labelled training data exists, although performance can remain uneven.
Streaming recognition systems must produce partial results with low latency, while offline transcription can exploit the entire recording and therefore use future context to improve accuracy. The continuing challenge is to achieve reliable recognition across languages, accents, environments and conversational styles while controlling computational cost and preserving privacy. Speech Recognition consequently remains both a mature engineering discipline and an active research area involving signal processing, linguistics, machine learning and human-computer interaction.
