I want to share a concrete example of using an LLM to automate a manual process in my workflow. Not chatbot stuff. An actual pipeline step that used to require a human sitting with two audio tracks for hours.
The problem
I build live speech-to-speech translation. To measure latency, I need to know which phrase in the source audio corresponds to which phrase in the translated audio, so I can measure the time gap between them. This alignment used to be done by hand. A person listens to both tracks, matches up the phrases, and logs timestamps. For a 6-minute session that’s easily an afternoon of work.
The hard part isn’t the math. It’s the alignment. Languages reorder things. German puts verbs at the end. Arabic restructures sentences. A Spanish phrase at position 3 might map to an English phrase at position 7.
Where the LLM fits in
This is exactly the kind of thing LLMs are good at. They understand semantic equivalence across languages and handle reordering naturally. So I replaced the manual step with an LLM call:
- Force-align both audio tracks to get per-word timestamps (automated, no LLM needed)
- Number every word in both transcripts and send them to the LLM
- LLM returns matched phrase pairs with word indices
- Compute the time gap for each pair using the timestamps from step 1
What used to take hours now takes a couple of minutes. No human in the loop.
The general pattern
The reason I’m sharing this is that the pattern generalizes. If you have a workflow step where a human reads two things and figures out how they correspond, an LLM can probably do it. The key is that I’m not asking it for a judgment call or creative output. I’m asking it to do structured alignment, a well-constrained task where it’s reliable.
The LLM only handles the one step that actually needs language understanding. Everything else (force alignment, timestamp extraction, aggregation) is regular code.
Full methodology: Automating ear-voice span
Code: VoiceFrom/live-s2st-eval
