Tortoise TTS v2: open-source multi-voice speech synthesis, slow but realistic
In one sentence James Betker releases Tortoise TTS 2.0, an open-source (Apache 2.0) text-to-speech program that imitates a voice from a few reference clips and targets realistic prosody and intonation, at the cost of very slow generation.
Tortoise is an open-source speech synthesis program written by James Betker alone, on his own hardware. Its stated priorities: speaking in many different voices and realistic prosody and intonation.
A few reference recordings are enough to imitate a voice: the guide recommends at least 3 clips of about 10 seconds. From the clips the model picks up tone, speed, timbre and even the recording conditions.
The name ("tortoise") is an admission: the model is extremely slow. According to the author, a K80 GPU takes about 2 minutes for a medium-length sentence. That makes it mostly suited to non-real-time uses.
Companies
Independent
Tools
Tortoise TTS
Tags
Sources