What is the best practice for extra long audio?

[dingkwang]

For long audio (e.g.,> 20 min), what is the best practice? What do you think of the following method?

• Split into 30s chunks, with 10s overlap.
• Get a transcript for each chunk.
• Manually or use LLM to connect to a full transcript.

[FireRedTeam]

1. use VAD to split long wav, each part shorter than 60s (FireRedASR-AED) or 30s (FireRedASR-LLM)
2. use FireRedASR to do ASR

[dingkwang]

But splitting wav may cause incomplete sentences or lost context.

[Albertchamberlain]

这些所有问题最终都会归结到 LLM的记忆和上下文长度上，算是一种tradeoff吧