[KongchangAI]
BenchmarkLibriTTS-R

LibriTTS-R

A large-scale, high-quality speech synthesis dataset derived from LibriTTS with improved audio quality through speech restoration techniques. It is commonly used for training and evaluating TTS models.

Timeline (last 90 days)

Oct 19

Evaluated with ±2 standard deviation requests on held-out LibriTTS-R speakers, SteerTTS achieves Spearman correlation coefficients between 0.87 and 0.97 for loudness, median pitch, H1–H2 and spectral tilt; however, ρ is only 0.46–0.79 for pauses, voicing ratio and CPPS, and CPPS leaks into pitch and level.

Unverified50%

All Facts (1)

Unverified

Evaluated with ±2 standard deviation requests on held-out LibriTTS-R speakers, SteerTTS achieves Spearman correlation coefficients between 0.87 and 0.97 for loudness, median pitch, H1–H2 and spectral tilt; however, ρ is only 0.46–0.79 for pauses, voicing ratio and CPPS, and CPPS leaks into pitch and level.

50%

Source Articles

SteerTTS: A 23M-Parameter TTS Model with Hybrid Attention and Bidirectional Mamba-3 for Fine-Grained Prosody ControlOct 19