README.md

The STREAMIND challenge

Welcome to the STREAMIND Grand Challenge!

The objective of the challenge is to design and implement a complete streaming pipeline, bridging the gap between live audio reception and structured language model inference. Participants will build their solutions on top of Juturna, an open-source Python framework developed by Meetecho for real-time AI data pipeline prototyping.

Important dates

• Registration deadline: April 3, 2026
• Paper submission deadline: June 19, 2026
• Acceptance notification: July 17, 2026

To register, submit a form here.

Challenge overview

The core task of this challenge is the implementation of a pipeline that consumes live audio streams as inputs, and produces textual summaries and keywords as outputs.

Roughly speaking, such pipeline can be organized around five main processing steps:

1. Audio reception via WebRTC/RTP: a live audio stream is delivered as a Opus-encoded mono RTP stream on a dedicated port.
2. Incremental ASR transcription: the incoming audio stream is incrementally transcribed by an ASR node operating on short, consecutive and partially overlapping audio chunks of a set temporal length.
3. Novel chunk extraction: novel transcript text is isolated while content from previous audio chunks is discarded.
4. Window aggregation: transcription chunks are accumulated into context windows of 300 seconds.
5. Summarization: aggregated windows are processed to produce structured textual outputs.
6. Transmission: summary objects are stored locally on the filesystem and transmitted to a destination endpoint through POST requests.

Evaluation metrics

All submitted chunks for an audio source will be assigned a composite score.

Chunk $i$, associated with the reference summary $ref_i$, contains the generated summary $gen_i$ and the keyword list $[k_1, k_2, k_3]_i$. Its summary score $S_i$ can then be computed using the LLM-as-a-judge technique, and normalized between 0 and 30.

$$ S_i = \text{Judge}(ref_i, gen_i, [k_1, k_2, k_3]_i) : 0 <= S_i <= 30 $$

The final chunk score $C_i$ is obtained by subtracting the latency associated with the chunk $i$ from its summary score.

$$ C_{i} = S_i - {\text{latency}_i} $$

The final score for an audio source is the average of all scores of its chunks.

References and Resources

• Juturna framework: https://github.com/meetecho/juturna
• Janus WebRTC server: https://github.com/meetecho/janus-gateway