Text-to-Speech and AI Voiceover

Text-to-speech (TTS) is the generation of spoken audio from written text. AI voiceover is its most common creative use: narration for videos, explainers, ads, audiobooks and games produced by a synthetic voice instead of, or alongside, a recorded one. Neural TTS crossed the threshold of natural-sounding speech some years ago. The 2026 frontier is control and speed: directing emotion and delivery, cloning a specific voice from seconds of audio, responding fast enough for live conversation, and covering a hundred languages. The same capabilities make consent and disclosure central to any responsible use. For the broader field, see speech and voice AI.

How Modern TTS Works

Older systems stitched together recorded fragments or used statistical models with a robotic edge. Current systems are neural and come in two main designs:

• Codec language models. Audio is compressed into discrete tokens by a neural codec, and a transformer predicts those tokens from text, much as a language model predicts words. Microsoft's 2023 VALL-E paper framed TTS this way and showed zero-shot voice cloning from a three-second sample. Many 2025–26 models, open and commercial, follow the pattern.
• Diffusion and flow models. These generate audio features by refining noise, then turn them into a waveform with a separate decoder. They are the other common design.

On top of either design, models add prosody control, meaning how a line is delivered. That ranges from the W3C's SSML markup for pauses and emphasis to inline emotion and style tags and plain-language direction ("warm, slightly amused"). Voice cloning conditions the model on reference audio of a particular speaker. Voice design creates a new voice from a text description.

The Landscape as of October 2026

Commercial APIs lead on quality, languages and tooling:

• ElevenLabs released Eleven v4 and v4 Turbo on September 28, 2026, covering more than 90 languages. Turbo is aimed at real-time use, at about 100 milliseconds of model latency.
• Cartesia made Sonic-3.6 generally available on August 27, 2026, with 44 languages. It focuses on low-latency voice agents.
• Google made gemini-3.8-flash-tts generally available on September 22, 2026. It offers voice design and "voice replication" gated by consent verification.
• OpenAI deprecated tts-1, tts-1-hd and its dated gpt-4o-mini-tts snapshots on October 1, 2026, with shutdown on January 6, 2027. It points developers to its realtime models instead.

The two specialists are compared in ElevenLabs vs Cartesia.

Open-weight models have closed much of the gap, though their licenses differ sharply:

• Boson AI's Higgs TTS 3 (about 4 billion parameters) covers 102 languages according to its model card. It ships under a research and non-commercial license with a "Creator Use Grant" that permits monetized creator content with credit to Boson AI. Production and API use need a commercial agreement. It is compared with the market leader in ElevenLabs vs Higgs Audio.
• Resemble AI's Chatterbox-Flash (May 2026, MIT) is a 0.5-billion-parameter English model.
• Nari Labs' Dia2 (November 2025, Apache-2.0) comes in 1B and 2B sizes and is built for dialogue.
• Kokoro-82M (Apache-2.0) is a very small model whose v1.0 offered 8 languages and 54 voices. It is popular for on-device use.

Real-Time Versus Studio Quality

TTS is used in two very different regimes. Conversational use, in voice agents, phone systems and game characters, needs the first audio within a few hundred milliseconds and tolerates small imperfections. That is why vendors advertise latency figures and "turbo" or "flash" variants. Studio use, in narration, audiobooks and ads, can wait seconds per line but demands consistent voice, correct pronunciation of names, and delivery that matches the picture. Most production voiceover work is studio work, where the practical concerns are different:

• Line-by-line generation. Generate each line separately, so a revised script regenerates only the lines that changed. That saves cost and keeps approved reads untouched.
• Fitting to picture. Check that each line fits the time available before the next shot or line, and adjust speaking rate or wording rather than letting narration run over a cut.
• Pronunciation. Product names, places and people's names need phonetic hints or custom dictionaries.
• Verification. Running the generated audio back through speech recognition catches skipped or garbled words before they ship.
• Mixing. Narration sits over music and natural sound, which means ducking and consistent loudness.

Consent, Rights and Disclosure

A cloned voice is a person's identity, and synthetic speech is a well-documented tool of fraud and deepfakes. The law is catching up:

• Tennessee's ELVIS Act (signed in March 2024, effective July 1, 2024) added voice to the state's protections against unauthorized commercial use of a person's likeness.
• The US NO FAKES Act was reintroduced in Congress in April 2025. It had not passed as of mid-2026.
• The EU AI Act's Article 50 requires machine-readable marking of synthetic audio from August 2026. Providers of systems already on the market have until December 2, 2026 for the marking and detection duties.

Vendors increasingly build consent into the product, as Google's consent-verified voice replication shows. For creators, the safe practice is simple. Clone only voices you have written permission to use, keep the consent record, disclose synthetic narration where platforms require it (see AI content labeling), and treat a voice model of a real person as sensitive data.

Where It's Headed

TTS is merging with translation and lip-sync into AI dubbing, where one performance becomes many languages in the speaker's own voice. It is also merging into speech-to-speech models that respond in real time without a separate text step. For video creators the near-term change is practical. Narration has become a draft that can be revised as easily as the script, with a human recording reserved for the lines that matter most. That shift touches digital humans, localization and the economics of generative audio alike.