Kugel 3

Kugel 3 is a high-fidelity, low-latency text-to-speech (TTS) model developed by KugelAudio, designed for real-time applications such as voice agents and content narration. It serves as the canonical production model for the KugelAudio platform, offering enhanced support for over 30 languages with a specialized focus on European languages including German, French, Spanish, and Polish. The model is optimized for sub-60ms first-packet latency and supports both zero-shot voice cloning and real-time audio streaming.

The architecture of the Kugel model family utilizes an Autoregressive (AR) + Diffusion framework, a design that enables fine-grained control over speech synthesis. This hybrid approach allows the model to handle complex linguistic tasks, such as the accurate pronunciation of email addresses, phone numbers, and technical terminology, with high precision. Kugel 3 is built upon research involving the YODAS2 dataset, having been trained on approximately 200,000 hours of highly pre-processed European speech data.

Key Capabilities

Kugel 3 features advanced zero-shot voice cloning, which is capable of replicating a speaker's timbre and prosody from a reference recording as short as 5 to 30 seconds. It also supports International Phonetic Alphabet (IPA) notation for custom pronunciations and provides built-in text normalization to ensure natural flow across different languages and dialects. The model is further distinguished by its ability to maintain high stability across a wide range of emotional expressions and speaking styles.

Control and Prompting

The model supports various inline tags to guide the generation process. Users can insert <break time="...ms"/> to introduce natural pauses, <spell> for character-by-character spelling, and <prosody rate="..."> to adjust the playback speed. Additionally, the cfg_scale parameter allows for control over the expressiveness of the synthesized voice; higher values typically increase the emotive quality of the output, while lower values provide more stable and neutral narration.