Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(realtime): accept bounded partial VAD overrun

LauraGPT980765099dea
diff --git a/funasr/bin/realtime_ws.py b/funasr/bin/realtime_ws.py
index 02728b2c5..5659b7cc1 100644
--- a/funasr/bin/realtime_ws.py
+++ b/funasr/bin/realtime_ws.py
@@ -1099,7 +1099,8 @@ class RealtimeASRSession:
and recent_eligible
and recent_observations >= 2
and int(recent_start_ms) == int(seg[0])
- and 0 <= tail_gap_ms <= max(100, decode_chunk_ms * 2)
+ and -max(100, decode_chunk_ms) <= tail_gap_ms
+ and tail_gap_ms <= max(100, decode_chunk_ms * 2)
and recent_partial
and not recent_hallucinated
):
diff --git a/tests/test_realtime_ws_service.py b/tests/test_realtime_ws_service.py
index a3686ff90..6d52b98b4 100644
--- a/tests/test_realtime_ws_service.py
+++ b/tests/test_realtime_ws_service.py
@@ -569,6 +569,48 @@ def test_long_segment_recovers_from_single_character_final():
assert text == partial
+def test_long_segment_accepts_partial_slightly_past_vad_end():
+ module = load_service_module()
+ partial = (
+ "哎,不好意思,上午好,高雄交通比较混乱,影响人有点。"
+ "哎,哎,胡总,各位长官,大家好。"
+ )
+ session = make_reported_long_segment_reconciliation_session(
+ module, partial, partial_end_ms=126976
+ )
+ session.last_partial_start_ms = 114430
+ session.segment_best_partial_start_ms = 114430
+
+ text = session._reconcile_completed_segment_text(
+ "哎",
+ [114430, 126740],
+ decode_succeeded=True,
+ )
+
+ assert text == partial
+
+
+def test_long_segment_rejects_partial_far_past_vad_end():
+ module = load_service_module()
+ partial = (
+ "哎,不好意思,上午好,高雄交通比较混乱,影响人有点。"
+ "哎,哎,胡总,各位长官,大家好。"
+ )
+ session = make_reported_long_segment_reconciliation_session(
+ module, partial, partial_end_ms=128800
+ )
+ session.last_partial_start_ms = 114430
+ session.segment_best_partial_start_ms = 114430
+
+ text = session._reconcile_completed_segment_text(
+ "哎",
+ [114430, 126740],
+ decode_succeeded=True,
+ )
+
+ assert text == "哎"
+
+
def test_long_segment_recovers_when_an_early_partial_correction_hides_tail_regression():
module = load_service_module()
partial = (