ASR Chunk Spacing
Preserve word boundaries when merging multi-chunk ASR output.
I’m testing long-form audio with Cohere Transcribe, and the transcript looks fine except at the points where vLLM moves to the next audio chunk. For example, one chunk ends with "Hello, this" and the next starts with "is vLLM", but the final response contains "Hello, thisis vLLM".
I’m seeing the same boundary issue with Qwen3 ASR in non-streaming transcription and translation. I need vLLM’s streaming and non-streaming speech responses to join clean per-chunk text consistently; with Cohere I can reproduce it in both modes, whether I leave the language unset or explicitly select English. I also tried Whisper, which looks fine. Longer transcripts contain several joined boundaries such as "superstar.A founding", "server.But I guess", and "mode.Putting this on".
Could you take a look? My guess is that a space is being lost when chunks from languages like English are joined. Do not add boundary spaces to Chinese or Japanese output, even when language is omitted. Find out what’s causing this and fix it.