How can I keep the Whisper model loaded in memory across multiple transcriptions, instead of reloading it for each audio file?

[YuXing404]

I’m building a tool that needs to transcribe multiple audio files one after another.

What’s the recommended way to keep the Whisper model loaded in memory across multiple transcriptions, so I don’t have to reload it from disk each time?

[officiallyutso]

You can load the model once using whisper_init_from_file() and reuse the returned whisper_context pointer across multiple calls to whisper_full().

struct whisper_context *ctx = whisper_init_from_file("models/ggml-base.en.bin");

// Reuse `ctx` for multiple audio files
whisper_full(ctx, params, audio1);
whisper_full(ctx, params, audio2);
// ...

whisper_free(ctx);

Just make sure your whisper_full_params is reset appropriately for each run. Also, avoid sharing the same ctx between threads unless protected.