StreamAudio-2M: Large-Scale Streaming Audio for AI Training

2.28 million unique audio clips are organized into streams for training audio-language models and agents. The dataset is structured into six task-specific subsets, including streams for audio understanding and real-time automatic speech recognition. It was created by zhifeixie and last updated on June 3, 2026.

Use Cases

• Train audio-language models based on streams of audio turns sharing a unified schema.
• Develop audio-understanding agents based on montages of clips with captions and question-answering data.
• Build real-time automatic speech recognition systems based on streams of ASR clips from sources like CommonVoice.
• Fine-tune models for multimodal tasks based on the structured, sequential nature of the audio streams.

Strengths

• Contains approximately 2.28 million unique audio clips.
• Organized into six distinct task subsets, including Stream_Audio_Understanding with 90,738 rows.
• Structured as streams, a sequence of audio turns, which is a distinctive format for sequential learning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the overall dataset is unknown, which may limit suitability assessment.
• Freshness should be verified as the last update timestamp is in the future (2026-06-03).

Provenance

Source

Aggregated from open data sources including AudioSet, FMA, CommonVoice, and GigaSpeech.

Collection Method

Clips are organized into streams (sequences of audio turns).

Freshness

Last updated 2026-06-03 13:18:23