Hume AI Launches Open-Source TTS TADA

🔑 Enhanced Key Takeaways

• •TADA achieves a real-time factor (RTF) of 0.09, enabling on-device deployment on mobile phones and edge devices without cloud dependency.[1]
• •TADA demonstrates zero content hallucinations across over 1,000 test samples due to its synchronized text-audio tokenization.[2]
• •Available models include a 1B parameter English version and a 3B parameter multilingual version on Hugging Face.[2]
• •TADA provides free transcripts alongside audio output with no additional latency.[2]

🛠️ Technical Deep Dive

• •TADA uses a novel text-acoustic dual alignment tokenization that synchronizes text and audio one-to-one at 2–3 frames (tokens) per second of audio, compared to 12.5–75 tokens per second in other LLM-based TTS systems.[1]
• •Architecture incorporates a diffusion head with flow matching sampling, converging at 4-10 steps per LLM decoding step, adding 50-75% per-token latency overhead but achieving overall RTF speedup due to low frame rate.[3]
• •Prompt processing uses a transcription model and wav2vec 2.0 Large-based aligner, with processed prompts cacheable for reuse; slightly higher peak memory but scales better for long outputs.[3]
• •In text-speech mode, TADA outperforms Spirit-LM-7B on tSC benchmark despite smaller size, fewer decoding steps, and continuous acoustic output versus Spirit-LM's 50Hz semantic tokens.[3]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline