The landscape of generative audio is shifting toward efficiency. A new open-source contender, Kani-TTS-2, has been released by the team at nineninesix.ai. This model marks a departure from heavy, compute-expensive TTS systems. Instead, it treats audio as a language, delivering high-fidelity speech synthesis with a remarkably small footprint. # Kani-TTS-2 follows the ‘Audio-as-Language’ philosophy. The model does not use traditional mel-spectrogram pipelines. Instead, it converts raw audio into discrete tokens using a neural codec. The system relies on a two-stage process: 1. Language Backbone: The model is built on LiquidAI’s LFM2 (350M) architecture. This backbone generates ‘audio intent’ by predicting the next audio tokens. Because LFM (Liquid Foundation Models) are designed for efficiency, they provide a faster alternative to standard transformers. 2. Neural Codec: It uses the NVIDIA NanoCodec to turn those tokens into 22kHz waveforms. By using this architecture, the model captures human-like prosody—the rhythm and intonation of speech—without the ‘robotic’ artifacts found in older TTS systems. ## Training at Warp Speed The training metrics for Kani-TTS-2 are a masterclass in optimization. The English model was trained on 10,000 hours of high-quality speech data. While that scale is impressive, the speed of training is the real story. The research team trained the model in only 6 hours using a cluster of 8 NVIDIA H100 GPUs. This proves that massive datasets no longer require weeks of compute time when paired with efficient architectures like LFM2. ## Zero-Shot Voice Cloning The standout feature for developers is zero-shot voice cloning. Unlike traditional models that require fine-tuning for new voices, Kani-TTS-2 uses speaker embeddings: - How it works: You provide a short reference audio clip. - The result: The model extracts the unique characteristics of that voice and applies them to the generated text instantly. ## Edge-Ready Performance From a deployment perspective, the model is highly accessible: | Specification | Value | |—————|——-| | Parameters | 400M (0.4B) | | Real-Time Factor (RTF) | 0.2 (10s audio in ~2s) | | VRAM Requirement | Only 3GB | | Compatible Hardware | RTX 3060, 4050, etc. | | License | Apache 2.0 (commercial-ready) | ## Why This Matters Kani-TTS-2 represents a significant shift in the TTS landscape: 1. Democratization: Running on consumer GPUs means developers no longer need expensive cloud APIs for production-quality TTS. 2. Local-First: Privacy-sensitive applications can now run entirely on-device. 3. Speed: The 0.2 RTF makes real-time interactive voice applications feasible. 4. AI Tools & Frameworks: Apache 2.0 licensing means commercial integration is straightforward. Kani-TTS-2 is available on Hugging Face in both [English (EN)https://huggingface.co/nineninesix/kani-tts-2-en){rel=“nofollow”} and [Portuguese (PT)https://huggingface.co/nineninesix/kani-tts-2-pt){rel=“nofollow”} versions.