StreamAudio-2M

# StreamAudio-2M Large-scale streaming-audio dataset for audio-LLM / audio-agent training. Each row is a **stream**: a sequence of audio turns sharing one unified schema. ~2.28M unique audio clips are organised into six task subsets. ## Subsets | Subset | Rows | Description | |---|---|---| | `Stream_Audio_Understanding` | 90,738 | Montages of audio-understanding clips (AudioSet / FMA): captions, choice & open QA | | `Real_time_ASR` | 28,109 | Streams of ASR clips (CommonVoice / GigaSpeech / LibriSpeech / VoxPopuli); `assistant` = transcript | | `Stream_translation` | 11,712 | Streams of EN→ZH speech-translation clips (CoVoST2); `assistant` = Chinese translation | | `Voice_chatting` | 196,064 | Multi-round spoken dialogues; each turn has spoken `user` + `assistant` | | `Proactive_respond` | 4,519 | Indoor/outdoor proactive-response streams (ElevenLabs / AudioX) | | `Environment_awared_audio_agent` | 50,035 | Same-scene audio montages (90–180 s) | ## Schema Top level: `id`, `stream_scene_type`, `stream_audio_category`, `num_turns`, `total_duration_s`, `turns` `stream_scene_type` / `stream_audio_category` hold the common value when all turns agree, else `"mixed"`. Each item of `turns`: `user`, `assistant`, `emotion`, `keywords`, `scene_type`, `audio_category`, `topic_type`, `audio_path`, `source`, `audio_stats` `audio_stats` = `{duration_s, sample_rate, channels, mean, std, abs_mean, rms_db, peak_db, crest_factor, dynamic_range_db, zero_crossing_rate}`. `scene_type` ∈ {Driving, Traffic, Home Smart, Cooking, Fitness, Travel, Office, others}. ## Audio Audio is shipped as uncompressed tar shards under `audio_tars/`. Download all shards and extract them into the dataset root to recreate the `audio/` tree; the `audio_path` field in every row is the relative path inside that tree, e.g. `audio/asr/0012/asr_0012345.wav`. ```bash cat audio_tars/*.tar | tar -xf - # or extract each shard individually ```