admin-valentin/WhisperLiveKit
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix simul-kv audio trim bug, add 1.7B v2 alignment heads

Browse source
This commit is contained in:
Quentin Fuxa 2026-03-15 20:45:00 +01:00
parent 7b08ea74ab
commit a540a5fd10
2 changed files with 3297 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3292
scripts/alignment_heads_qwen3_asr_1.7B_v2.json Normal file
View file

File diff suppressed because it is too large Load diff

7
whisperlivekit/qwen3_simul_kv.py
View file

@ -41,8 +41,8 @@ class Qwen3SimulKVConfig:
border_fraction: float = 0.20 border_fraction: float = 0.20
rewind_fraction: float = 0.12 rewind_fraction: float = 0.12
audio_min_len: float = 0.5 audio_min_len: float = 0.5
audio_max_len: float = 20.0 audio_max_len: float = 30.0
max_context_tokens: int = 30 max_context_tokens: int = 20
init_prompt: Optional[str] = None init_prompt: Optional[str] = None
max_alignment_heads: int = 10 max_alignment_heads: int = 10
@ -101,6 +101,9 @@ class Qwen3SimulKVState:
self.prompt_token_count = 0 self.prompt_token_count = 0
self.audio_token_count = 0 self.audio_token_count = 0
self.generated_token_ids = [] self.generated_token_ids = []
# Reset alignment tracking — old frame references are invalid
# after audio is trimmed from the front
self.last_attend_frame = -15
class Qwen3SimulKVASR: class Qwen3SimulKVASR: