LongSpeech

LongSpeech is a large-scale long-form audio understanding dataset with 100,000+ audio segments (~10 min each), designed to benchmark and train Audio LLMs on long-form speech. Presented at ICASSP 2026, it covers 8 tasks: ASR, translation, summarization, speaker counting, QA, and emotion analysis.

LongSpeech is a large-scale long-form audio understanding dataset with 100,000+ audio segments (~10 min each), designed to benchmark and train Audio LLMs on long-form speech. Presented at ICASSP 2026, it covers 8 tasks: ASR, translation, summarization, speaker counting, QA, and emotion analysis.

• 01100,000+ long-form audio segments (~10 minutes each) for benchmarking Audio LLMs
• 02Covers 8 tasks: ASR, translation, summarization, speaker counting, QA, emotion analysis
• 03Publicly released on Hugging Face with accompanying ICASSP 2026 paper