Part 3 of 4 · 2 min read
Exercises
Easy
Easy 1: Run it (30 minutes)
Run python -m pytest -q and python run_all.py in code/mini-voice.
Done when: you can explain each part of the 1.8-second gap.
Easy 2: Your budget (30 minutes)
Replace the stage latencies with ones you measured or got from a vendor's docs.
Done when: you report your own p50 and p95 gap.
Hint
vendor documentation often quotes time to first token and time to first audio; measure end to end anyway.
Easy 3: Shorten for voice (30 minutes)
Take five long chat replies from an agent you built and pass them through for_voice.
Done when: each is under 12 seconds spoken and still correct.
Medium
Medium 1: Smarter endpointing (3 hours)
Add a rule that waits longer after a trailing "and" or "um" and shorter after a complete sentence.
Done when: you show a better cut-in versus gap curve in the simulation.
Hint
log pause lengths from real conversations before choosing thresholds.
Medium 2: Mid-sentence correction (3 hours)
Handle "...oh wait, make that tomorrow" by keeping the booking and replacing the date.
Done when: five test utterances produce the right final booking.
Medium 3: Read-back confirmation (3 hours)
Add a confirmation step for amounts and dates before any irreversible tool call.
Done when: a test shows a misheard number is caught by the read-back.
Hint
spell numbers in groups when reading back.
Hard
Hard 1: Real calls (8 hours)
Instrument a real voice agent (a vendor realtime API or your own pipeline) and record ten calls. (I did not run this, so your numbers are the ground truth.)
Done when: you report gap p50 and p95, cut-in rate and interruption handling.
Hard 2: Telephony conditions (8 hours)
Test with phone-codec audio, background noise and packet loss. (I did not run this.)
Done when: you report word error rate on your domain terms versus clean audio.
Hint
telephony is typically 8 kHz audio, so models tuned on wideband audio may degrade.
Hard 3: Voice versus chat (8 hours)
Run the same task in voice and chat with five users. (I did not run this.)
Done when: you report completion time, errors and preference, and decide where voice should be used.