crosswk
a7e2dd6590
- 后端: ASR 性能优化与热词幻觉修复 (asr.py / main.py)，调整 config.example.yaml vllm_max_model_len - 补充 docs AI/ASR 服务提供商文档 - 同步客户端原生层(src-tauri)、overlay、recorder 等既有改动 - .gitignore 忽略 we-media/ 运营资料
109 lines
3.2 KiB
YAML
109 lines
3.2 KiB
YAML
# =============================================================================
|
|
# SayIt Configuration
|
|
# =============================================================================
|
|
# Secrets (API keys, passwords) → .env
|
|
# Everything else → this file
|
|
# =============================================================================
|
|
|
|
server:
|
|
host: "0.0.0.0"
|
|
port: 8000
|
|
cors_origins: # Allowed CORS origins
|
|
- "*" # Default: allow all. Restrict in production.
|
|
|
|
# HTTPS reverse proxy (when not using ALB/CDN)
|
|
gateway:
|
|
host: "0.0.0.0"
|
|
port: 8443
|
|
backend_target: "http://127.0.0.1:8000"
|
|
tls_key_file: "certs/dev.key"
|
|
tls_cert_file: "certs/dev.crt"
|
|
|
|
paths:
|
|
prompt_dir: "prompts"
|
|
releases_dir: "releases"
|
|
|
|
# --- Speech Recognition ---
|
|
asr:
|
|
engine: "qwen3" # qwen3 / firered
|
|
model: "Qwen/Qwen3-ASR-1.7B"
|
|
device: "cuda:0"
|
|
language: "Chinese" # Chinese / English / Cantonese
|
|
itn: true # Inverse text normalization
|
|
vllm_gpu_util: 0.5
|
|
vllm_max_model_len: 16384
|
|
max_concurrency: 1
|
|
batch_max_size: 32
|
|
batch_wait_ms: 60
|
|
firered: # Only used when engine = "firered"
|
|
model_dir: "pretrained_models"
|
|
|
|
vad:
|
|
model: "fsmn-vad"
|
|
max_single_segment_time: 30000 # ms
|
|
|
|
# --- LLM Text Polishing ---
|
|
# Define all available providers below, then choose which one to use
|
|
# for desktop and web_demo. API keys go in .env file.
|
|
llm:
|
|
desktop: "openai" # Provider for desktop client
|
|
web_demo: "groq" # Provider for web demo (empty = disabled)
|
|
|
|
providers:
|
|
azure:
|
|
deployment: "gpt-4o"
|
|
api_version: "2024-12-01-preview"
|
|
timeout_sec: 15
|
|
openai:
|
|
base_url: "https://api.openai.com"
|
|
model: "gpt-4o-mini"
|
|
timeout_sec: 15
|
|
groq:
|
|
base_url: "https://api.groq.com/openai"
|
|
model: "qwen/qwen3-32b"
|
|
timeout_sec: 15
|
|
ollama:
|
|
url: "http://127.0.0.1:11434/api/generate"
|
|
model: "qwen2.5:7b"
|
|
timeout_sec: 15
|
|
|
|
# --- Landing Page ---
|
|
public_site:
|
|
app_name: "SayIt"
|
|
headline: "随口说，出色写。"
|
|
subheadline: "用说话代替打字，AI 实时把口语变成可以直接用的书面表达。"
|
|
download_label: "下载 Windows 客户端"
|
|
download_platform: "win32"
|
|
download_arch: "x64"
|
|
|
|
# --- Browser Demo ---
|
|
web_demo:
|
|
enabled: true
|
|
max_duration_sec: 600 # Max recording length per session (sec)
|
|
max_concurrency_per_ip: 3
|
|
|
|
# --- Admin Dashboard ---
|
|
# Password in .env (SAYIT_ADMIN_PASSWORD)
|
|
admin:
|
|
enabled: true
|
|
username: "admin"
|
|
|
|
# --- Logging ---
|
|
logging:
|
|
level: "INFO" # DEBUG / INFO / WARNING / ERROR
|
|
debug_asr: false # Log raw ASR input/output
|
|
debug_llm: false # Log raw LLM input/output
|
|
file: "runtime/logs/sayit.log"
|
|
retention_days: 14
|
|
slow_asr_ms: 5000 # Warn when ASR takes longer
|
|
slow_llm_ms: 8000 # Warn when LLM takes longer
|
|
|
|
# --- Telemetry ---
|
|
telemetry:
|
|
enabled: true
|
|
db: "sqlite" # sqlite / postgresql
|
|
db_path: "runtime/telemetry/sayit.sqlite3"
|
|
collect_window_title: true
|
|
collect_exe_path: true
|
|
node_id: "single-node-01"
|