New Lightweight TTS Models Aim for High-Speed Voice Synthesis

A developer has released Lure-TTS and Lure-Turbo, two new text-to-speech (TTS) models designed for high-speed performance and granular control over vocal expression. According to reports from Reddit, the 600-million parameter models are engineered to run up to four times faster than real-time speech, targeting applications such as low-latency voice assistants.

The models were developed by fine-tuning the Qwen3-TTS architecture using advanced training techniques, including Reinforcement Learning from Verifiable Rewards (RLVR) and On-Policy Distillation (OPD). The developer noted that the primary objective was to create a smaller, more efficient model capable of running on consumer-grade hardware without the need for high-end GPUs.

Performance Benchmarks and Technical Specifications

In preliminary testing reported by the developer, Lure-TTS demonstrated a Real-Time Factor (RTF) of 0.14, compared to 0.195 for the 5B Fish S2 Pro. Despite the significant reduction in model size, the developer claims the model maintains comparable speaker similarity, citing an ECAPA score of 0.699.

The “Turbo” variant was achieved by optimizing the code predictor from 15 passes down to three, significantly increasing synthesis speed. The developer stated that the entire training process was completed for under $200, utilizing hardware no more powerful than an NVIDIA RTX PRO 6000.

To facilitate deployment, a dedicated engine has been released that supports CUDA (NVIDIA), MLX (Mac), and audio.cpp (CPU). The developer claims this engine can deliver the first audio chunk in 0.17 seconds—a significant reduction from the 10.6 seconds required by the standard Qwen-TTS reference code on an NVIDIA 4090.

Advanced Control and User Feedback

Lure-TTS introduces capabilities for “inline” control, allowing users to use tags to direct specific vocal behaviors. These include:

  • Events: Such as laughs, cries, sighs, and gulps.
  • Direction and Emotion: Allowing for specific tonal shifts.
  • Voice Cloning: Enabling the model to mimic specific speakers.

While the technical capabilities are significant, initial user feedback on community forums has been mixed. Some users noted that the synthesized audio can sound robotic, while others observed that the model occasionally struggles to prioritize emotional directions over the characteristics of the reference audio used for cloning.

The developer is reportedly working on improving the model’s ability to handle complex tags, such as (laugh: big) or (whispering), particularly during longer sequences of text. The project is currently available via GitHub, with the developer seeking community feedback on model outputs and potential research collaborations.