Gradium TTS is a low-latency speech synthesis model and API developed by Gradium, a commercial spin-off of the Kyutai AI research lab. Released in its upgraded form in June 2026, the model is architected specifically for real-time interactive applications, such as AI voice agents and live translation services. It prioritizes production reliability, focusing on minimizing Time to First Audio (TTFA) and maximizing pronunciation accuracy for complex alphanumeric strings, technical acronyms, and formatting.
The system is built on the Delayed Streams Modeling (DSM) architecture, which utilizes a Residual Quantized (RQ) Transformer. This approach represents audio signals as discrete tokens through Residual Vector Quantization (RVQ), allowing the model to process text and audio streams concurrently. This design achieves a median TTFA of approximately 155 ms, making it one of the lowest-latency streaming TTS solutions available at its release. In addition to its cloud API, Gradium provides an on-device variant known as Phonon, which utilizes a 100M parameter architecture optimized for local CPU execution.
Capabilities and Multilingual Support
The model supports high-fidelity synthesis across multiple languages, including English, French, German, Spanish, and Portuguese. It features integrated support for instant voice cloning, allowing developers to create custom voices from short audio samples, and provides pronunciation dictionaries to ensure accuracy for domain-specific terminology. The synthesis engine is designed to handle natural prosody and emotional nuances, making it suitable for conversational agents and gaming environments.
Developer Optimization Tips
For optimal performance and minimal latency, Gradium recommends several best practices for implementation:
- Sample Rate Matching: Maintain input at 24 kHz for recognition and output at 48 kHz PCM for synthesis to preserve fidelity without extra resampling.
- Semantic VAD: Utilize the integrated Semantic Voice Activity Detection to improve turn-taking by predicting user intent rather than relying solely on acoustic silence.
- Regional Endpoints: Use regional WebSocket endpoints (e.g., us.api.gradium.ai or eu.api.gradium.ai) to reduce round-trip time.
- In-Text Controls: Leverage in-text control tags and JSON configuration for fine-grained control over speech speed and rewrite rules.