Speech-to-text technology, commonly abbreviated STT and also called automatic speech recognition or ASR, is technology that converts spoken language captured as an audio signal into machine-readable text. A microphone first captures variations in air pressure produced by speech and converts them into an electrical or digital representation. Software then analyses the resulting audio, identifies linguistic patterns and estimates which sequence of words most likely corresponds to the speech.
Modern systems commonly use machine-learning models, particularly deep neural networks, to perform acoustic, phonetic and linguistic inference. The output can be displayed as ordinary text, entered into another application, used as subtitles, indexed for search, transformed into commands or processed by another artificial-intelligence system. Speech-to-text is therefore not simply a recording technology because its principal purpose is to infer linguistic content from sound rather than preserve the original sound itself.
Traditional speech-recognition systems separated the problem into several stages, including acoustic modelling, pronunciation modelling and language modelling. The acoustic model estimated relationships between audio features and speech sounds, the pronunciation model represented how words corresponded to sequences of phonemes, and the language model estimated which word sequences were linguistically plausible. Earlier systems often relied on hidden Markov models, Gaussian mixture models and statistical language models. Modern end-to-end neural systems can combine several of these functions within large learned models, reducing the need for manually designed intermediate representations. Transformer architectures, connectionist temporal classification and sequence-to-sequence approaches are among the techniques used in contemporary speech recognition.
The recognition process begins with sampling an acoustic waveform. An analogue microphone signal is converted into digital samples at a particular sampling frequency and represented with finite numerical precision. Speech-recognition systems may transform this signal into time-frequency representations such as spectrograms or other learned acoustic features. The system then analyses patterns associated with speech sounds, syllables, words and larger linguistic structures. Context is essential because many spoken sounds are ambiguous in isolation. The system may therefore use surrounding words, grammar, vocabulary and conversational context to determine whether a particular acoustic sequence is more likely to represent one phrase than another. This explains why recognition quality depends on both the audio signal and the language model.
Modern STT systems can handle continuous natural speech rather than requiring speakers to pronounce isolated words. They can incorporate punctuation, capitalisation, speaker diarisation and formatting as additional processing tasks. Speaker diarisation attempts to determine who spoke when, which is particularly useful for meetings, interviews and recorded conversations involving several people. Some systems can identify or adapt to particular accents, specialised vocabularies and domain terminology.
Custom vocabulary mechanisms can improve recognition of names, technical terms, medical terminology, product names or other words that occur infrequently in general training data. However, adaptation does not guarantee perfect recognition, particularly where acoustic conditions or vocabulary differ substantially from the data used to train the model.
Recognition accuracy is affected by many variables. Background noise, reverberation, microphone quality, distance from the speaker, overlapping speech, pronunciation, accent, speaking rate, code-switching and unusual vocabulary can all affect performance. A clean close-microphone recording generally provides more useful information than distant speech contaminated by machinery or conversation. Automatic gain control, noise suppression, beamforming and echo cancellation can improve the input before recognition. Multi-microphone systems can use spatial information to emphasise the desired speaker and suppress competing sounds. Modern models are often robust to some environmental variation, but robustness is not unlimited because severe signal corruption can remove acoustic information that no recognition algorithm can reliably reconstruct.
Speech-to-text has numerous practical applications. It enables dictation and voice typing, generates captions for video and live events, supports transcription of meetings and interviews, enables searchable audio archives, assists call-centre analysis and provides an input modality for voice-controlled software. Accessibility is another important application because speech recognition can allow people to enter text without conventional keyboard input and can support captioning for spoken content.
In conversational systems, STT frequently acts as the first processing stage: spoken input is transcribed, the resulting text is interpreted by another system, and a response may subsequently be generated and converted back into speech. STT can therefore form one component of a larger human-computer interaction pipeline.
Speech-to-text differs from simple keyword spotting, voice activity detection and speaker recognition. Keyword spotting detects whether particular words or phrases are present without necessarily transcribing all speech. Voice activity detection determines whether an audio segment contains speech rather than identifying its linguistic content. Speaker recognition attempts to determine or verify who is speaking rather than what is being said. Speech translation adds another stage by converting recognised speech into a different language. Text summarisation can operate after transcription to condense the resulting content. These distinctions matter because a system advertised as voice technology may perform one or several of these functions rather than full unrestricted transcription.
The principal limitations of speech-to-text arise from uncertainty, context dependence and the variability of human speech. A transcript can contain substitutions, omissions, insertions, incorrect punctuation or incorrect identification of specialist terminology. Errors can be particularly consequential in legal, medical, financial or safety-critical contexts, where human review may be required. Privacy is also important because cloud-based transcription can involve transmitting recordings or transcripts to external infrastructure, while local or on-device recognition can reduce some transmission requirements. Modern STT systems have become substantially more capable through large-scale neural modelling, but the underlying task remains probabilistic pattern recognition rather than a perfect conversion of sound into language.
