Add whisper-stream-pcm: stdin/pipe PCM streaming with VAD support

[rmorse]

Summary

• Adds whisper-stream-pcm, a new example binary for streaming raw PCM audio via stdin/pipe for real-time transcription — no WAV headers or SDL2 required
• Supports both simple (energy-based) and Silero VAD for speech segmentation
• Accepts 32-bit float or 16-bit signed integer PCM (16kHz, mono)

Usage

whisper-stream-pcm -m models/ggml-base.en.bin

With Silero VAD:

whisper-stream-pcm -m models/ggml-base.en.bin --vad --vad-model models/silero-vad.bin

Using FFMPEG to test/pipe in the data

ffmpeg -i audio.wav -f f32le -ar 16000 -ac 1 pipe:1 | whisper-stream-pcm -m models/ggml-base.en.bin --vad --vad-model models/silero-vad.bin

Changes

• examples/stream-pcm/ — new example binary (stream-pcm.cpp, CMakeLists.txt, README.md)
• examples/CMakeLists.txt — register stream-pcm example
• tests/test-stream-pcm.cpp — test for stdin/pipe PCM streaming
• tests/CMakeLists.txt — register test
• README.md — mention stream-pcm in examples list

Test plan

• whisper-stream-pcm builds on Linux/macOS/Windows
• Piping PCM audio via ffmpeg produces transcription output
• --vad (simple) and --vad --vad-model (Silero) both work correctly
• ctest passes test-stream-pcm

[ggerganov]

Thanks, this looks interesting. Would you mind rebasing on the latest master so we can run the CI

[rmorse]

@ggerganov done :) Thanks for considering this.

After double checking, there were some upstream warnings / patterns which I didn't change for stream-pcm but did find and fix a casting issue along the way.

[ggerganov]

Ok, let's see how the builds go.

I'm interested in adding such example. Need to find some time to play with this. Hopefully other people would provide some additional feedback in the meantime.

[rmorse]

I managed to a build a release against the last stable release and seemed to work well (on windows at least)

Just patching that CI issue now.