init: ECAPA-TDNN voiceprint embedding service

.gitignore

@@ -0,0 +1,6 @@
+venv/
+logs/
+__pycache__/
+*.pyc
+*.pid
+*.log

ahserver

@@ -0,0 +1 @@
+Subproject commit 6165795aa742d7539c758ff4ef370d24f3caaee7

app/voiceprint_app.py

@@ -0,0 +1,12 @@
+def extract_embedding(audio_path):
+    signal = MODEL.load_audio(audio_path)
+    if isinstance(signal, tuple):
+        signal = signal[0]
+    signal = signal.unsqueeze(0).to(DEVICE)
+    with torch.no_grad():
+        emb = MODEL.encode_batch(signal)
+    return emb.squeeze().cpu().numpy().tolist()
+
+def verify_speakers(audio_path, ref_path):
+    score, pred = MODEL.verify_files(audio_path, ref_path)
+    return float(score), bool(pred)

appPublic

@@ -0,0 +1 @@
+Subproject commit 371fe768cb28525ade1cab87c2f51b8e118a8fb3

conf/config.json

@@ -0,0 +1,33 @@
+{
+    "device": "cuda:1",
+    "redis_url": "redis://127.0.0.1:6379",
+    "worker_cnt": 2,
+    "logger": {
+        "name": "voiceprint",
+        "levelname": "info",
+        "logfile": "$[workdir]$/logs/voiceprint.log"
+    },
+    "website": {
+        "paths": [
+            ["$[workdir]$/wwwroot", ""]
+        ],
+        "client_max_size": 10000,
+        "host": "0.0.0.0",
+        "port": 9087,
+        "coding": "utf-8",
+        "indexes": ["index.html"],
+        "startswiths": [
+            {"leading": "/api/voiceprint-submit", "registerfunction": "voiceprint_submit"},
+            {"leading": "/api/voiceprint-status", "registerfunction": "voiceprint_status"},
+            {"leading": "/api/voiceprint/extract", "registerfunction": "voiceprint_extract"},
+            {"leading": "/api/voiceprint/verify", "registerfunction": "voiceprint_verify"}
+        ],
+        "processors": [
+            [".tmpl", "tmpl"],
+            [".ui", "bui"],
+            [".dspy", "dspy"]
+        ],
+        "session_max_time": 3000,
+        "session_issue_time": 2500
+    }
+}

download_model.py

@@ -0,0 +1,24 @@
+#!/usr/bin/env python3
+"""Download ECAPA-TDNN model via ModelScope to /share/models/ecapa-tdnn/"""
+import os, sys, time
+
+MODEL_ID = 'iic/speech_ecapa-tdnn_sv_en_voxceleb_16k'
+TARGET = '/share/models/ecapa-tdnn'
+
+os.makedirs(TARGET, exist_ok=True)
+
+# Try ModelScope first
+for attempt in range(3):
+    try:
+        from modelscope import snapshot_download
+        print(f"Attempt {attempt+1}: ModelScope download...")
+        snapshot_download(MODEL_ID, local_dir=TARGET)
+        print("SUCCESS via ModelScope")
+        sys.exit(0)
+    except Exception as e:
+        print(f"ModelScope failed: {e}")
+        time.sleep(10)
+
+# Fallback: try HF mirror
+print("Trying HF mirror...")
+os.system(f'export HF_ENDPOINT=https://hf-mirror.com && python3 -c "from huggingface_hub import snapshot_download; snapshot_download(\'speechbrain/spkrec-ecapa-voxceleb\', local_dir=\'{TARGET}\')"')

longtasks

@@ -0,0 +1 @@
+Subproject commit 3497f30c96b4b8beee551a8f2c2144e3581edd73

sqlor

@@ -0,0 +1 @@
+Subproject commit a9a02eb45bf8b0f5fa17c859eb4478c93a42a0e8

start.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+mkdir -p logs
+setsid python3 app/voiceprint_app.py -p 9087 >> logs/voiceprint.log 2>&1 &
+echo $! > voiceprint.pid
+echo "Voiceprint PID=$(cat voiceprint.pid)"
+sleep 3
+curl -s http://localhost:9087/api/health && echo "" || echo "(checking...)"

stop.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+if [ -f voiceprint.pid ]; then
+    kill $(cat voiceprint.pid) 2>/dev/null
+    rm -f voiceprint.pid
+    echo "Voiceprint stopped"
+else
+    pkill -f voiceprint_app.py 2>/dev/null && echo "Voiceprint stopped (via pkill)"
+fi

vp/__init__.py

@@ -0,0 +1,7 @@
+from ahserver.serverenv import ServerEnv
+from vp.engine import VoiceprintEngine
+
+
+def load_voiceprint():
+    env = ServerEnv()
+    env.voiceprint = VoiceprintEngine()

vp/engine.py

@@ -0,0 +1,63 @@
+"""ECAPA-TDNN Voiceprint Engine — loads model, processes extract/verify tasks."""
+import torch
+from longtasks.longtasks import LongTasks
+from appPublic.worker import awaitify
+from appPublic.jsonConfig import getConfig
+from appPublic.log import debug
+from speechbrain.inference.speaker import SpeakerRecognition
+import numpy as np
+
+
+class VoiceprintEngine(LongTasks):
+    def __init__(self):
+        self.config = getConfig()
+        super().__init__(self.config.redis_url, 'voiceprint', worker_cnt=self.config.worker_cnt)
+        self.load_model()
+
+    def load_model(self):
+        device = self.config.device
+        debug(f'loading ECAPA-TDNN on {device}...')
+        self.verifier = SpeakerRecognition.from_hparams(
+            source="speechbrain/spkrec-ecapa-voxceleb",
+            savedir="/share/models/ecapa-tdnn",
+            run_opts={"device": device}
+        )
+        debug('ECAPA-TDNN loaded')
+
+    async def process_task(self, payload, workerid=None):
+        task_type = payload.get('task_type', 'extract')
+        audio_file = payload.get('audio_file', '')
+        if not audio_file:
+            return {'status': 'FAILED', 'result': 'missing audio_file'}
+
+        if task_type == 'extract':
+            f = awaitify(self._extract)
+            return await f(audio_file)
+        elif task_type == 'verify':
+            ref_file = payload.get('reference_file', '')
+            if not ref_file:
+                return {'status': 'FAILED', 'result': 'missing reference_file'}
+            f = awaitify(self._verify)
+            return await f(audio_file, ref_file)
+        return {'status': 'FAILED', 'result': f'unknown task_type: {task_type}'}
+
+    def _extract(self, audio_path):
+        signal = self.verifier.load_audio(audio_path, 16000)
+        t = torch.tensor(signal).unsqueeze(0).to(self.config.device)
+        emb = self.verifier.encode_batch(t)
+        vec = emb.squeeze().cpu().numpy().tolist()
+        return {
+            'status': 'SUCCEEDED',
+            'embedding': vec,
+            'embedding_dim': len(vec),
+            'usage': {'audio_duration': round(len(signal) / 16000, 2)},
+        }
+
+    def _verify(self, audio_path, ref_path):
+        score, pred = self.verifier.verify_files(audio_path, ref_path)
+        return {
+            'status': 'SUCCEEDED',
+            'similarity': round(float(score), 4),
+            'is_same_speaker': bool(pred),
+            'usage': {},
+        }