Capabilities
Text-to-SpeechTTS
Also known as: speech synthesis, voice synthesis
AI that turns written text into natural-sounding spoken audio; the "voice" of assistants, audiobooks and accessibility tools.
Text-to-speech is AI that reads text out loud in a natural voice. It gives voice assistants their spoken replies, narrates audiobooks and articles, and helps people who cannot read a screen. Modern versions sound remarkably human, with natural rhythm and emotion, and can even mimic a specific person's voice, which is useful but also raises concerns about voice deepfakes.
Text-to-speech (TTS) generates spoken audio from written text. It is the output side of voice interfaces (the input side is speech-to-text).
Uses
- Voice assistants' spoken responses.
- Audiobooks, article narration, accessibility (screen readers).
- Voiceovers and characters in media.
How it improved
Older systems sounded robotic. Neural generative models learn to produce realistic waveforms with natural prosody (rhythm, stress, intonation), and voice cloning can reproduce a particular speaker from a short sample, which is powerful and also a misuse risk (voice deepfakes, scam calls).
Neural TTS synthesises speech from text, historically as a two-stage pipeline (text→acoustic features→vocoder) and increasingly end-to-end. The neural-audio era began with autoregressive raw-waveform generation (WaveNet; van den Oord et al. 2016) and progressed through faster parallel vocoders and, recently, LLM-style and diffusion-based approaches enabling high-fidelity, controllable, zero-shot voice cloning from short references.
Key dimensions: naturalness/prosody, expressiveness and emotion control, latency (critical for real-time multimodal voice agents), multilinguality, and speaker similarity for cloning. Evaluation leans on subjective mean opinion scores alongside objective intelligibility metrics. The same controllability that enables accessibility and product voices makes voice deepfakes and fraud feasible, motivating consent, provenance/watermarking and detection, and disclosure requirements for synthetic voices under emerging regulation.