42 lines
1.5 KiB
Bash
42 lines
1.5 KiB
Bash
# Local model assets; the backend never downloads models at startup.
|
||
MODEL_DIR=models
|
||
# Set this if CAM++ is not at a model_manifest.json path.
|
||
# CAM_MODEL_PATH=iic/speech_campplus_sv_zh-cn_16k-common
|
||
|
||
# FunASR realtime engine
|
||
FUNASR_ASR_MODEL=paraformer-zh-streaming
|
||
FUNASR_VAD_MODEL=fsmn-vad
|
||
FUNASR_DEVICE=cuda:0
|
||
FUNASR_VAD_DEVICE=cpu
|
||
FUNASR_CHUNK_SIZE=0,10,5
|
||
FUNASR_ENCODER_LOOK_BACK=4
|
||
FUNASR_DECODER_LOOK_BACK=1
|
||
FUNASR_VAD_CHUNK_MS=200
|
||
FUNASR_MAX_SEGMENT_SEC=30
|
||
# CAM++ 段内说话人追踪沿用 FunASR 的重叠短窗方案：1.5 秒窗、0.75 秒步长。
|
||
FUNASR_SPEAKER_WINDOW_MS=1500
|
||
FUNASR_SPEAKER_HOP_MS=750
|
||
# CAM++ 每批最多计算多少个窗口，避免很长 turn 一次性占用过多显存。
|
||
FUNASR_SPEAKER_BATCH_SIZE=16
|
||
# FunASR HybridSpeakerTracker history and stable identity matching defaults.
|
||
FUNASR_SPEAKER_HISTORY_CHUNKS=128
|
||
FUNASR_SPEAKER_MATCH_THRESHOLD=0.6
|
||
FUNASR_SPEAKER_MERGE_THRESHOLD=0.78
|
||
# FunASR's default speaker identity limit; increase this value for larger meetings.
|
||
FUNASR_MAX_SPEAKERS=15
|
||
# 小于该时长的相邻说话人段会合并，避免短窗噪声把一句话切得过碎。
|
||
FUNASR_SPEAKER_MIN_SEGMENT_MS=3000
|
||
|
||
# CAM++ is required and started by the backend launcher.
|
||
AUXILIARY_SERVICE_URL=http://127.0.0.1:8010
|
||
|
||
# Standalone frontend and browser-facing backend origin.
|
||
FRONTEND_HOST=127.0.0.1
|
||
FRONTEND_PORT=8080
|
||
FRONTEND_ORIGIN=http://127.0.0.1:8080
|
||
BACKEND_PUBLIC_URL=http://127.0.0.1:8082
|
||
|
||
WEB_HOST=0.0.0.0
|
||
WEB_PORT=8082
|
||
WEB_DISPLAY_HOST=127.0.0.1
|