nari-labs/dia2

Builds on the Kyutai Mimi codec to generate dialogue with speaker conditioning, enabling natural back-and-forth conversations by accepting audio context as input. Supports incremental generation from partial text without waiting for complete input, with 1B/2B model variants optimized for CUDA inference using bfloat16 precision and optional CUDA graph acceleration. Audio conditioning via Whisper transcription allows stable voice output when prefixed with speaker examples, supporting up to 2 minutes of English generation with word-level timestamps.

1,100 stars.