Text-to-speech, or TTS, converts written text into spoken audio. It is the output-facing counterpart to Speech recognition within Conversational AI systems, and has been used for decades in accessibility tools, phone systems, and navigation devices before becoming a mainstream generative-AI application in the 2020s.
History
Early TTS systems relied on concatenative synthesis, stitching together small pre-recorded segments of human speech, or formant synthesis, which generated speech from rules describing vocal-tract acoustics; both produced recognizably robotic output. A major shift came with the application of Deep learning to audio generation: DeepMind's WaveNet, published in 2016, generated raw audio waveforms sample by sample using a convolutional architecture and produced markedly more natural speech than prior methods, though at high computational cost. Subsequent systems such as Tacotron and later transformer-based models sped up generation while preserving much of the naturalness gain, and by the early 2020s neural TTS had become the default approach across the industry.
Modern systems
Contemporary TTS models typically separate the task into predicting an intermediate acoustic representation from text and then converting that representation into a waveform with a vocoder, though newer end-to-end systems increasingly collapse this into a single model, often built on Diffusion model or autoregressive Transformer (architecture) architectures. Companies such as ElevenLabs popularized highly expressive, low-latency TTS aimed at content creators, audiobook narration, and dubbing, offering control over emotion, pacing, and accent. TTS is also a core component of voice assistants, real-time translation tools, and accessibility software for visually impaired users, and it underlies AI-generated audio features such as NotebookLM's podcast-style summaries.
Voice cloning and quality
The most consequential recent advance is few-shot Voice cloning, where a model reproduces a specific person's voice from a short reference recording rather than a generic synthetic voice. As synthesized voices approached human indistinguishability, evaluation shifted from simple intelligibility tests toward perceptual naturalness scores and, increasingly, toward the ability to detect that audio is synthetic at all, a concern shared with the broader Deepfake literature.
Applications and concerns
TTS is used across audiobooks, customer-service phone systems, video narration, gaming, and assistive technology, and it has become a building block of larger Multimodal AI systems that can both listen and speak. Its rapid improvement has also raised concerns about voice-based fraud, impersonation in political disinformation, and the displacement of voice-acting work, prompting some companies to adopt AI watermarking or provenance labeling for synthetic audio and some jurisdictions to regulate the use of a person's voice without consent.