AuroraTTS-100K Releases: An Extremely Lightweight English TTS Model
AuroraAI-Research has released AuroraTTS-100K, an English text-to-speech model featuring under 100K parameters that produces mostly intelligible speech. When quantized to INT8, the model file size shrinks to just 127KB. This release represents a remarkable technical milestone for edge AI and local inference, demonstrating that functional text-to-speech is achievable with extreme efficiency. Its tiny footprint opens up new possibilities for embedding speech synthesis into resource-constrained devices. The model achieves an impressive inference speed of 72x real-time on a CPU and 1120x real-time on a GPU. However, it struggles with complex words and long numbers, which can sometimes result in slurred and unintelligible pronunciation.
## BACKGROUND
Text-to-speech (TTS) models typically require millions of parameters and substantial computational resources to synthesize natural-sounding audio. Model quantization is a technique that reduces the memory and computational costs of running inference by representing weights with low-precision data types like 8-bit integers.