seeding

transcription-svc/tests/whisper.rs

@@ -0,0 +1,38 @@
+//! End-to-end test of the real whisper.cpp engine: a recorded speech fixture is
+//! decoded by ffmpeg and transcribed. Ignored by default because it needs a
+//! downloaded model (`scripts/fetch-model.sh`) and ffmpeg on PATH; run with:
+//!   cargo test -p transcription-svc --test whisper -- --ignored
+
+use std::path::PathBuf;
+
+use transcription_svc::{whisper::WhisperTranscriber, Transcriber};
+
+fn manifest(rel: &str) -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join(rel)
+}
+
+#[tokio::test]
+#[ignore = "needs a downloaded model and ffmpeg"]
+async fn transcribes_real_speech() {
+    let model = std::env::var("WHISPER_MODEL")
+        .map(PathBuf::from)
+        .unwrap_or_else(|_| manifest("models/ggml-tiny.en.bin"));
+    assert!(
+        model.exists(),
+        "model not found at {model:?} — run scripts/fetch-model.sh"
+    );
+
+    let transcriber = WhisperTranscriber::new(model.to_str().unwrap()).expect("load model");
+    let audio = std::fs::read(manifest("tests/fixtures/speech.wav")).expect("read fixture");
+
+    let text = transcriber
+        .transcribe(&audio, "audio/wav")
+        .await
+        .expect("transcribe")
+        .to_lowercase();
+
+    assert!(
+        text.contains("fox"),
+        "expected 'fox' in transcript, got: {text:?}"
+    );
+}