All nine audio runs produced replies. The delay still makes this local combination a poor candidate to move straight into calls.

Sam wanted to know whether Pipecat could help with the pauses that had made an earlier phone conversation awkward. I ran a smaller offline check first. After the initial run, the median time from the end of the prerecorded input to the first generated audio was 2.9 seconds.

That is already a noticeable wait, and it excludes the real phone path.

What the clock measured

The input was a complete prerecorded question. Whisper transcribed it, a local Qwen model generated a response through Pipecat's Ollama service, and a sentence-buffered Kokoro voice produced audio. The timer stopped when the first audio samples were generated.

It did not stop when a caller heard a reply. There was no microphone capture, voice activity detection, turn detection, real-time transport, speaker playback, interruption handling or telephony in this measurement.

The three synthetic questions asked for the clinic phone number, a cancellation rule and an appointment booking. Each ran three times. The booking answer appropriately said it could not confirm or book appointments, although transcription rendered Nikki's name as Nicki.

The results

These are observations from a shared local machine, not a controlled ranking of voice frameworks. The speech recognition model, language model, sentence buffering and speech synthesis all contribute. Earlier component diagnostics also encountered severe memory pressure, which is another reason not to turn these trials into a clean performance comparison.

What I would change before calling again

I would first measure the individual stages under a stable machine load, then examine how much response text the speech step waits for. That would make the next change testable instead of assuming the framework name explains the delay.

After that, a separate real-time test needs to cover when the system decides a caller has finished speaking, how quickly audio reaches the caller and what happens when the caller interrupts. Those behaviors can make or break a conversation even if the language model itself is quick.

The verdict is more testing before calls. Pipecat did produce replies through the local pipeline; this run does not show that it fixed the earlier phone delay, nor does it establish that Pipecat caused it. No live clinic route was changed.

Test record

These findings reflect the October 2026 pilot. Read the public results and selected test evidence. Return to all five bonus reports.