🤖 AI Summary
The newly released *tiny-voca* is an innovative text-to-speech (TTS) model featuring just 4.5 million parameters and capable of generating speech in two voices: F2 (female) and M2 (male). It has been trained from scratch on approximately 18 hours of synthetic speech, enabling it to operate at an impressive speed of around 70 times real-time on standard CPUs. This low-cost solution, which was developed for less than $10 in GPU expenses, represents a significant leap for TTS technology, particularly for resource-constrained applications where efficiency is paramount.
What makes *tiny-voca* noteworthy for the AI/ML community is its architecture. It employs a transformer-based design that leverages explicit duration and prosody prediction, alongside a source-filter synthesis approach for waveform generation, bypassing the need for a neural vocoder. The model has demonstrated competitive performance with a word error rate (WER) comparable to larger systems, with each voice achieving decent rankings on TTS leaderboards. This development not only broadens accessibility to TTS technology but also paves the way for future innovations in voice clarity and personalization, particularly as improvements are planned for more mature voice options. The accessibility of the code, weights, and detailed datasets under an Apache-2.0 license further encourages experimentation and adaptation within the community.
Loading comments...
login to comment
loading comments...
no comments yet