[Feat] Streaming TTS word-level timestamps via forced aligner (#4034)

Signed-off-by: wjinxu <1299461899@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
authored and GitHub committed 2026-06-22 04:51:45 +00:00
1 parent a60242092a
commit ad07f98b73
17 files changed
+2162 -20

No files matched your search

@@ -0,0 +1,19 @@
# Reference deploy knobs for Qwen/Qwen3-ForcedAligner-0.6B, the shared forced
# aligner for streaming TTS word timestamps. Pass with --forced-aligner-config;
# --forced-aligner overrides the model path without editing this file.
#
# Run as LLM(runner="pooling") with pooling_task="token_classify". Prompt tokens,
# chat template and word segmentation are Qwen-specific and live in
# qwen3_force_align_processor (not configurable here).
forced_aligner:
model: Qwen/Qwen3-ForcedAligner-0.6B
runner: pooling
architecture: Qwen3ASRForcedAlignerForTokenClassification
pooling_task: token_classify
gpu_memory_utilization: 0.30
dtype: bfloat16
max_model_len: 4096
trust_remote_code: true