Text-to-speech technology, commonly abbreviated TTS, is technology that converts written or machine-readable text into an audio signal representing spoken language. A TTS system analyses input text, determines how words should be pronounced and interpreted, and generates speech that can be played through a speaker or headphones. Modern systems can produce highly natural voices using neural-network-based synthesis, whereas earlier systems often relied on concatenating recorded speech units or generating speech through explicit acoustic models and signal-processing techniques. TTS is used in screen readers, navigation systems, accessibility tools, telecommunications, virtual assistants, audiobooks, customer-service systems, games, education and many other applications. Its fundamental purpose is to turn symbolic linguistic information into an acoustic representation that humans can hear and understand.
TTS traditionally involves several linguistic processing stages. Text normalisation converts written forms such as numbers, dates, abbreviations and symbols into representations suitable for pronunciation. Grapheme-to-phoneme conversion determines likely speech sounds for words, while linguistic analysis can identify syntactic and prosodic information. A pronunciation lexicon may provide explicit pronunciations for known words, and rules or statistical models can handle unfamiliar forms. The system then generates acoustic information such as spectral characteristics, timing and fundamental frequency before a waveform is produced. Modern neural architectures can combine many of these stages and learn mappings from text or linguistic representations directly to speech waveforms or intermediate acoustic representations.
Pronunciation is more complex than simply reading letters aloud because written language contains ambiguity. A word such as record can have different pronunciations depending on whether it is functioning as a noun or verb. Numbers can represent quantities, dates, telephone numbers, measurements or addresses, and abbreviations can have multiple expansions. Text normalisation attempts to determine the intended spoken form from context. Proper names and newly created words can be particularly difficult because their pronunciation may not be predictable from spelling. TTS systems therefore require linguistic knowledge or learned representations capable of handling context, pronunciation variation and the conventions of the target language.
Prosody is a major factor in natural-sounding speech. It encompasses features such as rhythm, stress, intonation, pauses, duration and changes in pitch. Two systems can pronounce exactly the same sequence of words while sounding very different because their prosody differs. Human speakers naturally vary pitch and timing according to sentence structure, emphasis, emotion and conversational context. Modern neural TTS systems attempt to reproduce these characteristics by learning statistical relationships between text, linguistic structure and recorded speech. More advanced systems can provide controllable speaking styles, voices, speeds and emotional or expressive characteristics, although the degree of control and realism varies considerably between systems.
Modern neural TTS commonly uses architectures capable of learning complex mappings between linguistic input and acoustic or waveform output. Earlier neural systems often generated intermediate representations such as mel-spectrograms and then used a neural vocoder to synthesise the final waveform. More recent architectures can perform increasingly direct waveform generation or use large pretrained models. Neural methods have substantially improved naturalness, pronunciation and expressive variation compared with many older synthesis approaches. However, naturalness is not synonymous with correctness: a voice can sound convincing while pronouncing a specialised word incorrectly or interpreting ambiguous text incorrectly.
TTS has important accessibility applications. Screen readers use speech synthesis to read digital interfaces, documents and web pages aloud, helping users who cannot conveniently read visual text. TTS can also assist people with reading difficulties, provide spoken navigation instructions and make digital information available in auditory form. In communication devices, speech synthesis can generate output for users who cannot produce natural speech themselves. Because TTS produces speech from symbolic text, it can also generate entirely new utterances without requiring a separate recording for every sentence, which makes it substantially more flexible than systems based exclusively on prerecorded messages.
TTS differs from voice recording, voice conversion and speech generation in important ways. A recording reproduces speech that someone has previously produced, whereas TTS synthesises speech from textual input. Voice conversion transforms an existing speaker's voice characteristics into another target voice while retaining substantially the same linguistic content. TTS can generate novel sentences that were never recorded as complete utterances. Voice cloning technologies can use recordings of a particular speaker to construct a synthetic voice resembling that speaker, raising questions about consent, impersonation and misuse. These capabilities demonstrate that speech synthesis has both useful applications and security implications.
TTS quality depends on language coverage, pronunciation dictionaries, training data, model architecture, voice characteristics and input text. Errors can arise from ambiguous punctuation, unusual names, domain-specific vocabulary or incorrectly normalised numbers and symbols. Synthetic voices can also exhibit unnatural prosody, pronunciation errors or repetitive characteristics. Privacy and consent become especially important when systems reproduce identifiable human voices. Despite these limitations, modern TTS has progressed from highly mechanical-sounding synthesis towards speech that can be difficult to distinguish from natural human speech in some contexts. TTS is consequently a key interface technology linking textual information with auditory communication and is frequently paired with speech recognition to create complete spoken interaction systems.
