docs(stream): clarify VAD mode timestamps and CWD model path errors

[apollo-2006]

Added documentation for undocumented timestamp behavior in VAD mode and clarified CWD errors to prevent parser breaks," and hit submit

AI use: I found this while building against the library and used an AI assistant to help verify the relevant source (file/line refs above). The report and the documentation wording are mine, and I've checked every claim against the source myself.

[Copilot]

Pull request overview

This PR updates the whisper-stream example documentation to make its stdout output (especially in VAD / --step 0 mode) easier to parse reliably, and to clarify how model path resolution depends on the process working directory.

Changes:

• Documented the two distinct stdout output shapes for --step > 0 vs --step 0 (VAD mode), including the VAD block markers and timestamped segment lines.
• Clarified that -m/--model is resolved relative to the process CWD and that CWD mistakes can surface as the generic “failed to initialize whisper context” error.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[apollo-2006]

@danbev - you wrote #3065, so you're the right person to ask about the part I couldn't settle while writing this up.

When VAD is on, the timestamps that come back are relative to the filtered audio and get mapped back onto the original timeline. Reading the review thread on #3065, the mapping scales speech back across the original segment length rather than preserving where the silence actually sat. For a caller doing word-level alignment, is that difference meant to be treated as noise, or is preserving true silence positions something you'd want a follow-up for?

I hit this building a local voice assistant with a "Delphi" wake word on top of whisper-stream, so my failure mode was practical rather than theoretical: I was discarding timestamped lines as log noise until I realised the prefix was the data.

I'm a third-year CS student and I'm using this project mostly as an excuse to get better at reading code I didn't write. Thanks for #3065 either way.

[danbev]

@apollo-2006 The stream example uses vad_simple and not the VAD implementation from #3065.

For your question, please take a look at #3910 which may be relevant.

[apollo-2006]

Thanks, that's the part I had wrong. I was looking in the wrong place.

Checked stream.cpp: use_vad is just n_samples_step <= 0, and the only VAD is vad_simple in examples/common.cpp, which is an energy gate returning a bool. It never cuts silence or emits segments, and wparams.vad is never set, so nothing from #3065 runs here. My question didn't apply.

#3910 does answer what I was actually asking. Snapping a token that lands in removed silence to the nearer boundary, instead of interpolating across the gap, is the behaviour I meant. Merged, so nothing needed there.

One thing I did find: in --step 0, the ### Transcription N t0/t1 are wall-clock since stream start, while the [hh:mm:ss -->] prefixes are relative to the chunk passed to whisper_full. Two clocks, and a parser that mixes them gets it wrong. Want me to add that to this PR? Seems more useful than what's in there now.