luobangwang
/
test
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

调整speakerid出现时机

main
Bifang 2026-09-29 17:35:56 +08:00
parent d8e74d3293
commit fd621ab66e
1 changed files with 41 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
app/services/qwen3_websocket_asr.py
View File

@ -267,6 +267,8 @@ class ConnectionContext:
last_partial_raw_text: str = "" last_partial_raw_text: str = ""
last_partial_display_text: str = "" last_partial_display_text: str = ""
stable_partial_prefix: str = "" stable_partial_prefix: str = ""
pending_partial_revision_text: str = ""
pending_partial_revision_rounds: int = 0
segment_observed_text: str = "" segment_observed_text: str = ""
segment_observed_language: str = "" segment_observed_language: str = ""
best_partial_text: str = "" best_partial_text: str = ""
@ -1349,6 +1351,8 @@ class Qwen3ASRService:
ctx.last_partial_raw_text = "" ctx.last_partial_raw_text = ""
ctx.last_partial_display_text = "" ctx.last_partial_display_text = ""
ctx.stable_partial_prefix = "" ctx.stable_partial_prefix = ""
ctx.pending_partial_revision_text = ""
ctx.pending_partial_revision_rounds = 0
ctx.segment_observed_text = "" ctx.segment_observed_text = ""
ctx.segment_observed_language = "" ctx.segment_observed_language = ""
ctx.best_partial_text = "" ctx.best_partial_text = ""
@ -1696,6 +1700,8 @@ class Qwen3ASRService:
if not text: if not text:
ctx.last_partial_raw_text = "" ctx.last_partial_raw_text = ""
ctx.stable_partial_prefix = "" ctx.stable_partial_prefix = ""
ctx.pending_partial_revision_text = ""
ctx.pending_partial_revision_rounds = 0
return "" return ""
previous_raw = ctx.last_partial_raw_text previous_raw = ctx.last_partial_raw_text
@ -1719,9 +1725,30 @@ class Qwen3ASRService:
1, 1,
) )
if len(stable_prefix) - stable_common >= tolerated_divergence and previous_emitted: if len(stable_prefix) - stable_common >= tolerated_divergence and previous_emitted:
# A single shorter or divergent snapshot may be a transient model
# revision. Hold the visible text until the same correction repeats.
if text == ctx.pending_partial_revision_text:
ctx.pending_partial_revision_rounds += 1
else:
ctx.pending_partial_revision_text = text
ctx.pending_partial_revision_rounds = 1
if ctx.pending_partial_revision_rounds < 2:
ctx.last_partial_raw_text = text
return previous_emitted return previous_emitted
# The repeated revision is stable enough to replace the old prefix.
stable_prefix = text[: self._stable_prefix_cutoff(text, stable_common)]
ctx.pending_partial_revision_text = ""
ctx.pending_partial_revision_rounds = 0
else:
ctx.pending_partial_revision_text = ""
ctx.pending_partial_revision_rounds = 0
stable_prefix = text[:stable_common] stable_prefix = text[:stable_common]
stable_prefix = text[: self._stable_prefix_cutoff(text, len(stable_prefix))] stable_prefix = text[: self._stable_prefix_cutoff(text, len(stable_prefix))]
else:
ctx.pending_partial_revision_text = ""
ctx.pending_partial_revision_rounds = 0
ctx.stable_partial_prefix = stable_prefix ctx.stable_partial_prefix = stable_prefix
ctx.last_partial_raw_text = text ctx.last_partial_raw_text = text
@ -2789,6 +2816,13 @@ class Qwen3ASRService:
) )
observed = update_observed_text(ctx, current, current_language) observed = update_observed_text(ctx, current, current_language)
visible_observed = self._trim_previous_segment_overlap(ctx, observed) visible_observed = self._trim_previous_segment_overlap(ctx, observed)
if has_native_stream_snapshot:
# Qwen returns full-utterance snapshots; stabilize the
# displayed prefix while allowing repeated corrections.
visible_observed = self._stabilize_partial_text(
ctx,
visible_observed,
)
partial_display = self._clip_partial_text(visible_observed, ctx) partial_display = self._clip_partial_text(visible_observed, ctx)
if ( if (
partial_display partial_display
@ -2934,6 +2968,10 @@ class Qwen3ASRService:
emit_segment_start=False, emit_segment_start=False,
) )
# Drain pending diarization before the final event; the browser closes
# the socket on `end`, so later speaker updates would otherwise be lost.
await ctx.speaker_job_queue.join()
final_updates = self._recluster_confirmed_segments(ctx) final_updates = self._recluster_confirmed_segments(ctx)
for updated in final_updates: for updated in final_updates:
sentence_payload = self._build_tencent_sentence( sentence_payload = self._build_tencent_sentence(