From f27d844ed710541f0aab6298012f27cd337a5477 Mon Sep 17 00:00:00 2001 From: Bifang <915779419@qq.com> Date: Fri, 9 Oct 2026 15:40:36 +0800 Subject: [PATCH] =?UTF-8?q?vllm=E8=A7=A3=E7=A0=81=E6=97=A5=E5=BF=97?= =?UTF-8?q?=E5=A2=9E=E5=8A=A0operation=E3=80=81=E4=BC=9A=E8=AF=9Did?= =?UTF-8?q?=E3=80=81=E5=8E=9F=E7=94=9Fpartial=E9=98=9F=E5=88=97=E7=AD=89?= =?UTF-8?q?=E5=BE=85=E3=80=81=E9=94=81=E7=AD=89=E5=BE=85=E5=92=8C=E7=94=9F?= =?UTF-8?q?=E6=88=90=E8=80=97=E6=97=B6?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- app/services/qwen3_websocket_asr.py | 20 ++++++++++++++++++-- 1 file changed, 18 insertions(+), 2 deletions(-) diff --git a/app/services/qwen3_websocket_asr.py b/app/services/qwen3_websocket_asr.py index b5d065a..35d7817 100644 --- a/app/services/qwen3_websocket_asr.py +++ b/app/services/qwen3_websocket_asr.py @@ -3202,6 +3202,9 @@ class Qwen3ASRService: partial_decode_ms = (time.perf_counter() - partial_decode_started) * 1000.0 ctx.last_partial_decode_samples = int(ctx.segment_audio_buffer.size) current = self._sanitize_candidate_text(current) + observed_chars = 0 + stabilized_chars = 0 + preview_changed = False partial_display = "" partial_outcome = "empty" if current and not self._is_degenerate_repetition(current): @@ -3212,6 +3215,7 @@ class Qwen3ASRService: else self._update_segment_observed_text ) observed = update_observed_text(ctx, current, current_language) + observed_chars = len(observed) visible_observed = self._trim_previous_segment_overlap(ctx, observed) if has_native_stream_snapshot: # Qwen returns full-utterance snapshots; stabilize the @@ -3220,11 +3224,14 @@ class Qwen3ASRService: ctx, visible_observed, ) + stabilized_chars = len(visible_observed) + previous_preview = ctx.partial_preview_text partial_display = self._append_only_partial_preview( ctx, visible_observed, native_stream_snapshot=has_native_stream_snapshot, ) + preview_changed = partial_display != previous_preview partial_display = self._clip_partial_text(partial_display, ctx) if not partial_display: partial_outcome = "filtered_empty" @@ -3300,8 +3307,9 @@ class Qwen3ASRService: # path, without logging transcript contents or changing events. logger.info( "[%s] Realtime partial result: segment=%d source=%s outcome=%s " - "decode_ms=%.1f turn_audio_sec=%.2f window_audio_sec=%.2f " - "stream_chunk_id=%s candidate_chars=%d display_chars=%d", + "partial_resolve_ms=%.1f turn_audio_sec=%.2f window_audio_sec=%.2f " + "stream_chunk_id=%s window_rolled=%s candidate_chars=%d observed_chars=%d " + "stabilized_chars=%d preview_changed=%s display_chars=%d", task_id, ctx.segment_index, "native_stream" if has_native_stream_snapshot else "window_fallback", @@ -3310,7 +3318,15 @@ class Qwen3ASRService: len(ctx.segment_audio_buffer) / 16000.0, len(ctx.stream_window_buffer) / 16000.0, getattr(ctx.realtime_stream_state, "chunk_count", None), + getattr( + getattr(ctx.realtime_stream_state, "internal_state", None), + "window_rolled", + None, + ), len(current), + observed_chars, + stabilized_chars, + preview_changed, len(partial_display), )