Text-to-Speech (TTS)

Text-to-Speech (TTS) est la synthèse vocale convertissant texte en parole audio. Bond qualitatif dramatique depuis 2017 (Tacotron, WaveNet) puis 2023+ avec neural TTS atteignant qualité quasi-indistinguable de voix humaines.

Services et modèles leaders 2024-2025 : (1) **ElevenLabs** — leader qualité naturelle, voice cloning, multilingue, expressivité ; (2) **OpenAI TTS** (tts-1, tts-1-hd) — voices variées (Alloy, Echo, Fable, Onyx, Nova, Shimmer) ; (3) **Google Cloud Text-to-Speech** — Chirp 3 voices, Wavenet ; (4) **Azure AI Speech** — Neural TTS 400+ voices, Custom Neural Voice ; (5) **AWS Polly** — Neural Engine, Generative engine ; (6) **PlayHT** — voice cloning ; (7) **Cartesia Sonic** — ultra low latency ; (8) **Murf.ai** ; (9) **Speechify** ; (10) **Open source** : Coqui XTTS, Bark (Suno), StyleTTS 2, F5-TTS, Kokoro TTS.

Features : (1) **Voice variety** — multiple voices par language ; (2) **Voice cloning** — clone voice from few seconds samples (controversial copyright/deepfake concerns) ; (3) **Emotion control** — sad, happy, angry, surprised ; (4) **Multilingual** — single model parlant 30+ langues ; (5) **SSML** (Speech Synthesis Markup Language) — control pace, pitch, emphasis, pauses, pronunciation ; (6) **Streaming** TTS — low-latency for real-time conversation ; (7) **Realtime API** (OpenAI 2024) — full duplex voice conversation with GPT-4o.

Use cases : (1) accessibility (screen readers) ; (2) audiobook narration ; (3) voice assistants ; (4) IVR phone systems ; (5) e-learning ; (6) podcast generation ; (7) video voiceovers ; (8) gaming character voices ; (9) language learning ; (10) navigation systems.

Éthique : voice cloning facile soulève deepfake concerns (scams imitant familles, politiciens). Watermarking acoustique (SynthID Google, ElevenLabs Detector) tentative solutions. Réglementation émergente (EU AI Act, US state laws). Compétences AI-102, AIF-C01.