whisper : add stateless VAD detect + explicit state reset for streaming (#3677)

whisper_vad_detect_speech resets LSTM state on every call, which is
correct for batch processing but prevents temporal continuity when
calling per-chunk in a streaming loop.

Add whisper_vad_detect_speech_no_reset (skips buffer clear) and
whisper_vad_reset_state (explicit clear between utterances).
Existing whisper_vad_detect_speech is now a thin wrapper — zero
behavior change for current callers.

Co-authored-by: Claude Opus 4.6 (1M context) <redacted>

diff --git a/include/whisper.h b/include/whisper.h
index f4cc6bf7abdd96c0851ed506edd071daf6c0801f..b5dcdb2917a2fe76fbb164ecf8b321c0ed9600fb 100644 (file)
--- a/include/whisper.h
+++ b/include/whisper.h
@@ -695,6 +695,16 @@ extern "C" {
                            const float * samples,
                                    int   n_samples);
 
+    // Like whisper_vad_detect_speech, but does not reset LSTM state.
+    // Use for streaming: call whisper_vad_reset_state() between utterances.
+    WHISPER_API bool whisper_vad_detect_speech_no_reset(
+            struct whisper_vad_context * vctx,
+                           const float * samples,
+                                   int   n_samples);
+
+    // Reset LSTM hidden/cell states to zero.
+    WHISPER_API void whisper_vad_reset_state(struct whisper_vad_context * vctx);
+
     WHISPER_API int     whisper_vad_n_probs(struct whisper_vad_context * vctx);
     WHISPER_API float * whisper_vad_probs  (struct whisper_vad_context * vctx);
 

diff --git a/src/whisper.cpp b/src/whisper.cpp
index 86bfafeaad8d96bea78ad1e3cb17b4fcf026bae1..2f356da0f0670743fbe5bd97b7edda1a24cbab6f 100644 (file)
--- a/src/whisper.cpp
+++ b/src/whisper.cpp
@@ -5083,7 +5083,11 @@ struct whisper_vad_context * whisper_vad_init_with_params(
     return vctx;
 }
 
-bool whisper_vad_detect_speech(
+void whisper_vad_reset_state(whisper_vad_context * vctx) {
+    ggml_backend_buffer_clear(vctx->buffer, 0);
+}
+
+bool whisper_vad_detect_speech_no_reset(
         struct whisper_vad_context * vctx,
         const float * samples,
         int n_samples) {
@@ -5095,9 +5099,6 @@ bool whisper_vad_detect_speech(
     WHISPER_LOG_INFO("%s: detecting speech in %d samples\n", __func__, n_samples);
     WHISPER_LOG_INFO("%s: n_chunks: %d\n", __func__, n_chunks);
 
-    // Reset LSTM hidden/cell states
-    ggml_backend_buffer_clear(vctx->buffer, 0);
-
     vctx->probs.resize(n_chunks);
     WHISPER_LOG_INFO("%s: props size: %u\n", __func__, n_chunks);
 
@@ -5165,6 +5166,14 @@ bool whisper_vad_detect_speech(
     return true;
 }
 
+bool whisper_vad_detect_speech(
+        struct whisper_vad_context * vctx,
+        const float * samples,
+        int n_samples) {
+    whisper_vad_reset_state(vctx);
+    return whisper_vad_detect_speech_no_reset(vctx, samples, n_samples);
+}
+
 int whisper_vad_segments_n_segments(struct whisper_vad_segments * segments) {
     return segments->data.size();
 }