src-tauri/src/transcription/mod.rs

@@ -287,8 +287,22 @@ impl Transcriber for WhisperTranscriber {
 
 #[cfg(feature = "cpu-transcription")]
 fn available_threads() -> std::ffi::c_int {
+    // WA_WHISPER_THREADS: support/bench override — see cap rationale below.
+    if let Some(n) = std::env::var("WA_WHISPER_THREADS")
+        .ok()
+        .and_then(|s| s.parse().ok())
+    {
+        return n;
+    }
+    // Capped at 4 (whisper.cpp's own CLI default): ggml workers spin-wait at
+    // a per-graph-node barrier, so one worker scheduled on a slow core stalls
+    // every node while the rest burn CPU. With all 14 logical CPUs of a
+    // hybrid Meteor Lake (incl. 2 LP E-cores) this measured RTF ~22 — ~80x
+    // slower than the same build at 4 threads (RTF 0.28), regardless of
+    // compiler flags. 4 threads is also the CPU-frugal choice (NFR-RES-1):
+    // 1.07 vs 2.16 cpu-sec/audio-sec at 12 threads for near-equal wall time.
     std::thread::available_parallelism()
-        .map(|n| n.get() as std::ffi::c_int)
+        .map(|n| (n.get() as std::ffi::c_int).min(4))
         .unwrap_or(4)
 }