ElevenLabs

You submit text (optionally with inline tags like a whisper or a pause) and a neural text-to-speech model converts it directly into an audio waveform in the chosen voice, rather than stitching together pre-recorded fragments. Voice cloning works the same way in reverse: a short sample of someone's speech trains a voice profile the model can then use to speak new text. Dubbing and conversational agents chain this with speech recognition and translation models.