Kimi-Audio-7B-Instruct

Kimi-Audio-7B-Instruct is an open-source audio foundation model specialized for text-to-speech (TTS) and a broad range of audio processing tasks including speech recognition, audio question answering, captioning, emotion recognition, sound event classification, and end-to-end speech conversation. Hosted by Gigarouter as a managed API, it delivers production-ready inference without the overhead of model deployment.

Key capabilities

• Unified framework for diverse audio tasks within a single model
• State-of-the-art performance on numerous audio benchmarks (see technical report)
• Pre-trained on 13 million hours of speech, music, and environmental audio plus text data
• Hybrid audio input combining continuous acoustic tokens with discrete semantic tokens
• LLM core with parallel output heads for text and audio token generation
• Chunk-wise streaming detokenizer based on flow matching for low-latency audio synthesis

Ideal use cases

• Real-time TTS for conversational agents, voice assistants, and content narration
• High-quality speech-to-speech interaction with both text and audio output
• Multi-purpose audio understanding and generation in a single API call

Performance highlights

Kimi-Audio achieves state-of-the-art results across multiple audio benchmarks. For detailed comparisons, consult the technical report.

Additional resources

• Hugging Face model page
• GitHub repository
• Technical report (PDF)