feat: VAD

bun.lock

@@ -20,6 +20,7 @@
         "@aws-sdk/credential-providers": "^3.1020.0",
         "@azure/identity": "^4.13.1",
         "@commander-js/extra-typings": "^14.0.0",
+        "@ericedouard/vad-node-realtime": "^0.2.0",
         "@growthbook/growthbook": "^1.6.5",
         "@larksuite/channel": "^0.1.2",
         "@larksuiteoapi/node-sdk": "^1.60.0",
@@ -79,6 +80,7 @@
         "lru-cache": "^11.2.7",
         "marked": "^17.0.5",
         "node-edge-tts": "^1.2.10",
+        "onnxruntime-web": "^1.27.0",
         "p-map": "^7.0.4",
         "picomatch": "^4.0.4",
         "plist": "^3.1.0",
@@ -242,6 +244,8 @@
 
     "@emnapi/runtime": ["@emnapi/runtime@1.11.0", "", { "dependencies": { "tslib": "^2.4.0" } }, "sha512-55coeOFKHv1ywEcUXJtWU5f+Jr/W5tZDvZig8DLKSwUN1JpROQ4rk/SNOQiFWmaR/VKF4zuFyW1B8JduOSv6Pg=="],
 
+    "@ericedouard/vad-node-realtime": ["@ericedouard/vad-node-realtime@0.2.0", "", { "dependencies": { "onnxruntime-node": "^1.14.0" } }, "sha512-hVKup5v4x5ixqZ3TnyLQKsybE3GveVQZmUX3iaZT0mbSendCt7aMhqW1wxDm4fKFgOUryJyV2znaQU2Xyik9wQ=="],
+
     "@growthbook/growthbook": ["@growthbook/growthbook@1.6.5", "", { "dependencies": { "dom-mutator": "^0.6.0" } }, "sha512-mUaMsgeUTpRIUOTn33EUXHRK6j7pxBjwqH4WpQyq+pukjd1AIzWlEa6w7i6bInJUcweGgP2beXZmaP6b6UPn7A=="],
 
     "@grpc/grpc-js": ["@grpc/grpc-js@1.14.4", "", { "dependencies": { "@grpc/proto-loader": "^0.8.0", "@js-sdsl/ordered-map": "^4.4.2" } }, "sha512-k9Dj3DV/itK9D06Y8f190Qgop7/Ui+D0njFV3LHMPwPT75DpXLQohE9Wmz0QElrJnzsjB7KPWiKJbOl7IPDArQ=="],
@@ -496,6 +500,8 @@
 
     "accepts": ["accepts@2.0.0", "", { "dependencies": { "mime-types": "^3.0.0", "negotiator": "^1.0.0" } }, "sha512-5cvg6CtKwfgdmVqY1WIiXKc3Q1bkRqGLi+2W/6ao+6Y7gu/RCwRuAhGEzh5B4KlszSuTLgZYuqFqo5bImjNKng=="],
 
+    "adm-zip": ["adm-zip@0.5.17", "", {}, "sha512-+Ut8d9LLqwEvHHJl1+PIHqoyDxFgVN847JTVM3Izi3xHDWPE4UtzzXysMZQs64DMcrJfBeS/uoEP4AD3HQHnQQ=="],
+
     "agent-base": ["agent-base@8.0.0", "", {}, "sha512-QT8i0hCz6C/KQ+KTAbSNwCHDGdmUJl2tp2ZpNlGSWCfhUNVbYG2WLE3MdZGBAgXPV4GAvjGMxo+C1hroyxmZEg=="],
 
     "ajv": ["ajv@8.20.0", "", { "dependencies": { "fast-deep-equal": "^3.1.3", "fast-uri": "^3.0.1", "json-schema-traverse": "^1.0.0", "require-from-string": "^2.0.2" } }, "sha512-Thbli+OlOj+iMPYFBVBfJ3OmCAnaSyNn4M1vz9T6Gka5Jt9ba/HIR56joy65tY6kx/FCF5VXNB819Y7/GUrBGA=="],
@@ -608,8 +614,12 @@
 
     "default-browser-id": ["default-browser-id@5.0.1", "", {}, "sha512-x1VCxdX4t+8wVfd1so/9w+vQ4vx7lKd2Qp5tDRutErwmR85OgmfX7RlLRMWafRMY7hbEiXIbudNrjOAPa/hL8Q=="],
 
+    "define-data-property": ["define-data-property@1.1.4", "", { "dependencies": { "es-define-property": "^1.0.0", "es-errors": "^1.3.0", "gopd": "^1.0.1" } }, "sha512-rBMvIzlpA8v6E+SJZoo++HAYqsLrkg7MSfIinMPFhmkorw7X+dOXVJQs+QT69zGkzMyfDnIMN2Wid1+NbL3T+A=="],
+
     "define-lazy-prop": ["define-lazy-prop@3.0.0", "", {}, "sha512-N+MeXYoqr3pOgn8xfyRPREN7gHakLYjhsHhWGT3fWAiL4IkAt0iDw14QiiEm2bE30c5XX5q0FtAA3CK5f9/BUg=="],
 
+    "define-properties": ["define-properties@1.2.1", "", { "dependencies": { "define-data-property": "^1.0.1", "has-property-descriptors": "^1.0.0", "object-keys": "^1.1.1" } }, "sha512-8QmQKqEASLd5nx0U1B1okLElbUuuttJ/AnYmRXbbbGDWh6uS208EjD4Xqq/I9wK7u0v6O08XhTWnt5XtEbR6Dg=="],
+
     "delayed-stream": ["delayed-stream@1.0.0", "", {}, "sha512-ZySD7Nf91aLB0RxL4KGrKHBXl7Eds1DAmEdcoVawXnLD7SDhpNgtuII2aAkg7a7QS41jxPSZ17p4VdGnMHk3MQ=="],
 
     "depd": ["depd@2.0.0", "", {}, "sha512-g7nH6P6dyDioJogAAGprGpCtVImJhpPk/roCzdb3fIh61/s/nPsfR6onyMwkCAR/OlC3yBC0lESvUoQEAssIrw=="],
@@ -690,6 +700,8 @@
 
     "find-up": ["find-up@4.1.0", "", { "dependencies": { "locate-path": "^5.0.0", "path-exists": "^4.0.0" } }, "sha512-PpOwAdQ/YlXQ2vj8a3h8IipDuYRi3wceVQQGYWxNINccq40Anw7BlsEXCMbt1Zt+OLA6Fq9suIpIWD0OsnISlw=="],
 
+    "flatbuffers": ["flatbuffers@25.9.23", "", {}, "sha512-MI1qs7Lo4Syw0EOzUl0xjs2lsoeqFku44KpngfIduHBYvzm8h2+7K8YMQh1JtVVVrUvhLpNwqVi4DERegUJhPQ=="],
+
     "flora-colossus": ["flora-colossus@2.0.0", "", { "dependencies": { "debug": "^4.3.4", "fs-extra": "^10.1.0" } }, "sha512-dz4HxH6pOvbUzZpZ/yXhafjbR2I8cenK5xL0KtBFb7U2ADsR+OwXifnxZjij/pZWF775uSCMzWVd+jDik2H2IA=="],
 
     "follow-redirects": ["follow-redirects@1.16.0", "", {}, "sha512-y5rN/uOsadFT/JfYwhxRS5R7Qce+g3zG97+JrtFZlC9klX/W5hD7iiLzScI4nZqUS7DNUdhPgw4xI8W2LuXlUw=="],
@@ -728,6 +740,10 @@
 
     "glob": ["glob@13.0.6", "", { "dependencies": { "minimatch": "^10.2.2", "minipass": "^7.1.3", "path-scurry": "^2.0.2" } }, "sha512-Wjlyrolmm8uDpm/ogGyXZXb1Z+Ca2B8NbJwqBVg0axK9GbBeoS7yGV6vjXnYdGm6X53iehEuxxbyiKp8QmN4Vw=="],
 
+    "global-agent": ["global-agent@4.1.3", "", { "dependencies": { "globalthis": "^1.0.2", "matcher": "^4.0.0", "semver": "^7.3.5", "serialize-error": "^8.1.0" } }, "sha512-KUJEViiuFT3I97t+GYMikLPJS2Lfo/S2F+DQuBWzuzaMPnvt5yyZePzArx36fBzpGTxZjIpDbXLeySLgh+k76g=="],
+
+    "globalthis": ["globalthis@1.0.4", "", { "dependencies": { "define-properties": "^1.2.1", "gopd": "^1.0.1" } }, "sha512-DpLKbNU4WylpxJykQujfCcwYWiV/Jhm50Goo0wrVILAv5jOr9d+H+UR3PhSCD2rCCEIg0uc+G+muBTwD54JhDQ=="],
+
     "google-auth-library": ["google-auth-library@10.7.0", "", { "dependencies": { "base64-js": "^1.3.0", "ecdsa-sig-formatter": "^1.0.11", "gaxios": "^7.1.4", "gcp-metadata": "8.1.2", "google-logging-utils": "1.1.3", "jws": "^4.0.0" } }, "sha512-QpTAbNJ36TliZLx3TTtahR8HG0hN9RllL1e3FymOvQSIKK8JmgV58H924ub2wa2DsS3ANjjP1Aw1N+Ramc8hqQ=="],
 
     "google-logging-utils": ["google-logging-utils@1.1.3", "", {}, "sha512-eAmLkjDjAFCVXg7A1unxHsLf961m6y17QFqXqAXGj/gVkKFrEICfStRfwUlGNfeCEjNRa32JEWOUTlYXPyyKvA=="],
@@ -740,8 +756,12 @@
 
     "gtoken": ["gtoken@7.1.0", "", { "dependencies": { "gaxios": "^6.0.0", "jws": "^4.0.0" } }, "sha512-pCcEwRi+TKpMlxAQObHDQ56KawURgyAf6jtIY046fJ5tIv3zDe/LEIubckAO8fj6JnAxLdmWkUfNyulQ2iKdEw=="],
 
+    "guid-typescript": ["guid-typescript@1.0.9", "", {}, "sha512-Y8T4vYhEfwJOTbouREvG+3XDsjr8E3kIr7uf+JZ0BYloFsttiHU0WfvANVsR7TxNUJa/WpCnw/Ino/p+DeBhBQ=="],
+
     "has-flag": ["has-flag@5.0.1", "", {}, "sha512-CsNUt5x9LUdx6hnk/E2SZLsDyvfqANZSUq4+D3D8RzDJ2M+HDTIkF60ibS1vHaK55vzgiZw1bEPFG9yH7l33wA=="],
 
+    "has-property-descriptors": ["has-property-descriptors@1.0.2", "", { "dependencies": { "es-define-property": "^1.0.0" } }, "sha512-55JNKuIW+vq4Ke1BjOTjM2YctQIvCT7GFzHwmfZPGo5wnrgkid0YQtnAleFSqumZm4az3n2BS+erby5ipJdgrg=="],
+
     "has-symbols": ["has-symbols@1.1.0", "", {}, "sha512-1cDNdwJ2Jaohmb3sg4OmKaMBwuC48sYni5HUw2DvsC8LjGTLK9h+eb1X6RyuOHe4hT0ULCW68iomhjUoKUqlPQ=="],
 
     "has-tostringtag": ["has-tostringtag@1.0.2", "", { "dependencies": { "has-symbols": "^1.0.3" } }, "sha512-NqADB8VjPFLM2V0VvHUewwwsw0ZWBaIdgo+ieHtK3hasLz4qeCRjYcqfB6AQrBggRKppKF8L52/VqdVsO47Dlw=="],
@@ -852,6 +872,8 @@
 
     "marked": ["marked@17.0.6", "", { "bin": { "marked": "bin/marked.js" } }, "sha512-gB0gkNafnonOw0obSTEGZTT86IuhILt2Wfx0mWH/1Au83kybTayroZ/V6nS25mN7u8ASy+5fMhgB3XPNrOZdmA=="],
 
+    "matcher": ["matcher@4.0.0", "", { "dependencies": { "escape-string-regexp": "^4.0.0" } }, "sha512-S6x5wmcDmsDRRU/c2dkccDwQPXoFczc5+HpQ2lON8pnvHlnvHAHj5WlLVvw6n6vNyHuVugYrFohYxbS+pvFpKQ=="],
+
     "math-intrinsics": ["math-intrinsics@1.1.0", "", {}, "sha512-/IXtbwEk5HTPyEwyKX6hGkYXxM9nbj64B+ilVJnC/R6B0pH5G4V3b0pVbL7DBj4tkhBAppbQUlf6F6Xl9LHu1g=="],
 
     "media-typer": ["media-typer@1.1.0", "", {}, "sha512-aisnrDP4GNe06UcKFnV5bfMNPBUw4jsLGaWwWfnH3v02GnBuXX2MCVn5RbrWo0j3pczUilYblq7fQ7Nw2t5XKw=="],
@@ -898,12 +920,20 @@
 
     "object-inspect": ["object-inspect@1.13.4", "", {}, "sha512-W67iLl4J2EXEGTbfeHCffrjDfitvLANg0UlX3wFUUSTx92KXRFegMHUVgSqE+wvhAbi4WqjGg9czysTV2Epbew=="],
 
+    "object-keys": ["object-keys@1.1.1", "", {}, "sha512-NuAESUOUMrlIXOfHKzD6bpPu3tYt3xvjNdRIQ+FeT0lNb4K8WR70CaDxhuNguS2XG+GjkyMwOzsN5ZktImfhLA=="],
+
     "on-finished": ["on-finished@2.4.1", "", { "dependencies": { "ee-first": "1.1.1" } }, "sha512-oVlzkg3ENAhCk2zdv7IJwd/QUD4z2RxRwpkcGY8psCVcCYZNq4wYnVWALHM+brtuJjePWiYF/ClmuDr8Ch5+kg=="],
 
     "once": ["once@1.4.0", "", { "dependencies": { "wrappy": "1" } }, "sha512-lNaJgI+2Q5URQBkccEKHTQOPaXdUxnZZElQTZY0MFUAuaEqe1E+Nyvgdz/aIyNi6Z9MzO5dv1H8n58/GELp3+w=="],
 
     "onetime": ["onetime@5.1.2", "", { "dependencies": { "mimic-fn": "^2.1.0" } }, "sha512-kbpaSSGJTWdAY5KPVeMOKXSrPtr8C8C7wodJbcsd51jRnmD+GZu8Y0VoU6Dm5Z4vWr0Ig/1NKuWRKf7j5aaYSg=="],
 
+    "onnxruntime-common": ["onnxruntime-common@1.27.0", "", {}, "sha512-3KxL5wIVqa8Ex08jxSzncm9CMgw8CjOFyOQ7SxvG9o0cVLlhTNKXyIQuTbtX4tGPJEf73OER2xrjt4HJSBL4ow=="],
+
+    "onnxruntime-node": ["onnxruntime-node@1.27.0", "", { "dependencies": { "adm-zip": "^0.5.16", "global-agent": "^4.1.3", "onnxruntime-common": "1.27.0" }, "os": [ "linux", "win32", "darwin", ] }, "sha512-QEzGwrvNBgv4uPVdnbHsOGG4G6T96mdlcFI8aAKPjMU8wOPpVocPXb6k3QGkaZagVTv2G9Bnnbo6Z3JdXr1fQw=="],
+
+    "onnxruntime-web": ["onnxruntime-web@1.27.0", "", { "dependencies": { "flatbuffers": "^25.1.24", "guid-typescript": "^1.0.9", "long": "^5.2.3", "onnxruntime-common": "1.27.0", "platform": "^1.3.6", "protobufjs": "^7.2.4" } }, "sha512-ogDLsqIozHZwifPuN37OproAo0byX6t43/bP8GzeZWBWD6MOGExswFAx3up4NS/vvWBOg2u2PXomDt3rMmdQSg=="],
+
     "open": ["open@10.2.0", "", { "dependencies": { "default-browser": "^5.2.1", "define-lazy-prop": "^3.0.0", "is-inside-container": "^1.0.0", "wsl-utils": "^0.1.0" } }, "sha512-YgBpdJHPyQ2UE5x+hlSXcnejzAvD0b22U2OuAP+8OnlJT+PjWPxtgmGqKKc+RgTM63U9gN0YzrYc71R2WT/hTA=="],
 
     "opus-encdec": ["opus-encdec@0.1.1", "", {}, "sha512-TDzyGqYqrwn5UEUNaLsfLGu8Ma+HRNrgLYj7Vx5wfTnafAA21G6Bnm/qTIa3orQi/yZPZYmkdpO/gez4nfA1Rw=="],
@@ -942,6 +972,8 @@
 
     "pkce-challenge": ["pkce-challenge@5.0.1", "", {}, "sha512-wQ0b/W4Fr01qtpHlqSqspcj3EhBvimsdh0KlHhH8HRZnMsEa0ea2fTULOXOS9ccQr3om+GcGRk4e+isrZWV8qQ=="],
 
+    "platform": ["platform@1.3.6", "", {}, "sha512-fnWVljUchTro6RiCFvCXBbNhJc2NijN7oIQxbwsyL0buWJPG85v81ehlHI9fXrJsMNgTofEoWIQeClKpgxFLrg=="],
+
     "plist": ["plist@3.1.1", "", { "dependencies": { "@xmldom/xmldom": "^0.9.10", "base64-js": "^1.5.1", "xmlbuilder": "^15.1.1" } }, "sha512-ZIfcLJC+7E7FBFnDxm9MPmt7D+DidyQ26lewieO75AdhA2ayMtsJSES0iWzqJQbcVRSrTufQoy0DR94xHue0oA=="],
 
     "pngjs": ["pngjs@5.0.0", "", {}, "sha512-40QW5YalBNfQo5yRYmiw7Yz6TKKVr3h6970B2YE+3fQpsWcrbj1PzJgxeJ19DRQjhMbKPIuMY8rFaXc8moolVw=="],
@@ -1000,6 +1032,8 @@
 
     "send": ["send@1.2.1", "", { "dependencies": { "debug": "^4.4.3", "encodeurl": "^2.0.0", "escape-html": "^1.0.3", "etag": "^1.8.1", "fresh": "^2.0.0", "http-errors": "^2.0.1", "mime-types": "^3.0.2", "ms": "^2.1.3", "on-finished": "^2.4.1", "range-parser": "^1.2.1", "statuses": "^2.0.2" } }, "sha512-1gnZf7DFcoIcajTjTwjwuDjzuz4PPcY2StKPlsGAQ1+YH20IRVrBaXSWmdjowTJ6u8Rc01PoYOGHXfP1mYcZNQ=="],
 
+    "serialize-error": ["serialize-error@8.1.0", "", { "dependencies": { "type-fest": "^0.20.2" } }, "sha512-3NnuWfM6vBYoy5gZFvHiYsVbafvI9vZv/+jlIigFn4oP4zjNPK3LhcY0xSCgeb1a5L8jO71Mit9LlNoi2UfDDQ=="],
+
     "serve-static": ["serve-static@2.2.1", "", { "dependencies": { "encodeurl": "^2.0.0", "escape-html": "^1.0.3", "parseurl": "^1.3.3", "send": "^1.2.0" } }, "sha512-xRXBn0pPqQTVQiC8wyQrKs2MOlX24zQ0POGaj0kultvoOCstBQM5yvOhAVSUwOMjQtTvsPWoNCHfPGwaaQJhTw=="],
 
     "set-blocking": ["set-blocking@2.0.0", "", {}, "sha512-KiKBS8AnWGEyLzofFfmvKwpdPzqiy16LvQfK3yv/fVH7Bj13/wl3JSR1J+rfgRE9q7xUJK4qvgS8raSOeLUehw=="],
@@ -1392,6 +1426,8 @@
 
     "ink/wrap-ansi": ["wrap-ansi@9.0.2", "", { "dependencies": { "ansi-styles": "^6.2.1", "string-width": "^7.0.0", "strip-ansi": "^7.1.0" } }, "sha512-42AtmgqjV+X1VpdOfyTGOYRi0/zsoLqtXQckTmqTeybT+BDIbM/Guxo7x3pE2vtpr1ok6xRqM9OpBe+Jyoqyww=="],
 
+    "matcher/escape-string-regexp": ["escape-string-regexp@4.0.0", "", {}, "sha512-TtpcNJ3XAzx3Gq8sWRzJaVajRs0uVxA2YAkdb1jm2YkPz4G6egUFAyA3n5vtEIZefPk5Wa4UXbKuS5fKkJWdgA=="],
+
     "minipass-flush/minipass": ["minipass@3.3.6", "", { "dependencies": { "yallist": "^4.0.0" } }, "sha512-DxiNidxSEK+tHG6zOIklvNOwm3hvCrbUrdtzY74U6HKTJxvIDfOUL5W5P2Ghd3DTkhhKPYGqeNUIh5qcM4YBfw=="],
 
     "minipass-pipeline/minipass": ["minipass@3.3.6", "", { "dependencies": { "yallist": "^4.0.0" } }, "sha512-DxiNidxSEK+tHG6zOIklvNOwm3hvCrbUrdtzY74U6HKTJxvIDfOUL5W5P2Ghd3DTkhhKPYGqeNUIh5qcM4YBfw=="],
@@ -1410,6 +1446,8 @@
 
     "restore-cursor/signal-exit": ["signal-exit@3.0.7", "", {}, "sha512-wnD2ZE+l+SPC/uoS0vXeE9L1+0wuaMqKlfz9AMUo38JsyLSBWSFcHR1Rri62LZc12vLr1gb3jl7iwQhgwpAbGQ=="],
 
+    "serialize-error/type-fest": ["type-fest@0.20.2", "", {}, "sha512-Ne+eE4r0/iWnpAxD852z3A+N0Bt5RN//NjJwRd2VFHEmrywxf5vsZlh4R6lixl6B+wz/8d+maTSAkN1FIkI3LQ=="],
+
     "type-is/content-type": ["content-type@2.0.0", "", {}, "sha512-j/O/d7GcZCyNl7/hwZAb606rzqkyvaDctLmckbxLzHvFBzTJHuGEdodATcP3yIRoDrLHkIATJuvzbFlp/ki2cQ=="],
 
     "vscode-languageserver-protocol/vscode-jsonrpc": ["vscode-jsonrpc@9.0.0", "", {}, "sha512-+VvMmQPJhtvJ+8O+zu2JKIRiLxXF8NW7krWgyMGeOHrp4Cn23T5hc0v2LknNeopDOB70wghHAds7mKtcZ0I4Sg=="],

package.json

@@ -38,6 +38,7 @@
     "@aws-sdk/credential-providers": "^3.1020.0",
     "@azure/identity": "^4.13.1",
     "@commander-js/extra-typings": "^14.0.0",
+    "@ericedouard/vad-node-realtime": "^0.2.0",
     "@growthbook/growthbook": "^1.6.5",
     "@larksuite/channel": "^0.1.2",
     "@larksuiteoapi/node-sdk": "^1.60.0",
@@ -97,6 +98,7 @@
     "lru-cache": "^11.2.7",
     "marked": "^17.0.5",
     "node-edge-tts": "^1.2.10",
+    "onnxruntime-web": "^1.27.0",
     "p-map": "^7.0.4",
     "picomatch": "^4.0.4",
     "plist": "^3.1.0",

src/friend/FriendService.ts

@@ -19,6 +19,7 @@ import { getPrefs } from './prefs.js';
 import { stripForTts } from './text-utils.js';
 import { edgeTts, qwenTts, registerAudioFile } from './tts.js';
 import { splitSentences } from './text-utils.js';
+import { SileroVad } from './voice/vad-service.js';
 
 // ── Types ──────────────────────────────────────────────────────────────
 
@@ -69,14 +70,13 @@ class FriendService {
   private audioCapture: AudioCaptureProvider | null = null;
   /** Active STT connection (Anthropic/Doubao/Whisper) during capture */
   private sttConnection: { send: (chunk: Buffer) => void; finalize: () => Promise<void>; close: () => void } | null = null;
-  /** Periodic timer for server-side voice segmentation (~5s intervals) */
-  private captureTimer: ReturnType<typeof setInterval> | null = null;
-  private readonly SEGMENT_MS = 5000;
   /** STT provider/language cached for connection re-creation during segmentation */
   private captureProvider = '';
   private captureLanguage = '';
-  /** Guard against concurrent _flushSegment calls */
+  /** Guard against concurrent flush calls */
   private _flushing = false;
+  /** Silero VAD instance (real ML-based voice activity detection) */
+  private vadInstance: SileroVad | null = null;
 
   // ── React sync external store interface ──────────────────────────────
 
@@ -102,9 +102,22 @@ class FriendService {
     this.setState({ status: 'starting', lastError: undefined });
 
     try {
-      // Pre-warm cpal audio module (loaded on first use)
-      // The voice service lazy-loads audio-capture-napi, so the first
-      // capture will incur the ~1s dlopen penalty regardless.
+      // Pre-warm Silero VAD (loads ONNX model + onnxruntime-web WASM)
+      if (!this.vadInstance) {
+        const vad = new SileroVad({
+          onSpeechStart: () => {},
+          onSpeechEnd: (_audio) => {
+            this._flushVadSegment().catch((e) =>
+              console.error('[FriendService] VAD segment flush error:', e),
+            );
+          },
+        });
+        vad.init().then(() => {
+          this.vadInstance = vad;
+        }).catch((e) => {
+          console.warn('[FriendService] VAD init failed (non-fatal, voice capture falls back to F2-only):', e);
+        });
+      }
 
       this.setState({ status: 'running' });
     } catch (err) {
@@ -229,64 +242,56 @@ class FriendService {
       throw new Error('Native audio capture unavailable');
     }
 
-    // 4. Start periodic segmentation timer (voice call auto-send)
-    this._startSegmentTimer();
+    // 4. Start Silero VAD for real-time speech-end detection
+    // audio chunks flow to VAD via the arecord data callback (see loadAudioCapture)
+    if (this.vadInstance) {
+      try {
+        this.vadInstance.start();
+      } catch (e) {
+        console.warn('[FriendService] VAD start error (non-fatal):', e);
+      }
+    }
   }
 
   /**
-   * Periodically finalize the current STT segment and start a new one,
-   * sending the transcript to the CLI conversation in real-time.
+   * Flush the current STT segment and start a new one, triggered by
+   * Silero VAD's onSpeechEnd callback.
    *
-   * This replaces browser-based VAD with server-side timer segmentation.
-   * Without it, voice call text only appears when the user presses F2.
+   * The audio from the just-ended speech segment has already been sent
+   * to the STT connection. We swap to a fresh connection so the next
+   * speech segment starts clean.
    */
-  private async _flushSegment(): Promise<void> {
+  private async _flushVadSegment(): Promise<void> {
     if (this._flushing || !this.capturing) return;
     this._flushing = true;
     try {
       const oldConn = this.sttConnection;
       if (!oldConn) return;
 
-      // 1. Create a new STT connection for ongoing audio first
+      // 1. Create a new STT connection for ongoing audio
       const newConn = await this.startSttConnectionWithTimeout(
         this.captureProvider,
         this.captureLanguage,
       );
       this.sttConnection = newConn;
 
-      // 2. Finalize old connection (sends buffered audio to Groq)
+      // 2. Finalize old connection (sends buffered audio to STT provider)
       await oldConn.finalize().catch(() => {});
       oldConn.close();
 
-      // 3. Send any accumulated transcript to the CLI conversation
+      // 3. Send accumulated transcript to the CLI conversation
       const transcript = this.captureTranscripts.join('').trim();
       if (transcript) {
         this.captureTranscripts = [];
         this.sendText(transcript);
       }
     } catch (err) {
-      console.error('[FriendService] _flushSegment error:', err);
+      console.error('[FriendService] _flushVadSegment error:', err);
     } finally {
       this._flushing = false;
     }
   }
 
-  private _startSegmentTimer(): void {
-    this._clearSegmentTimer();
-    this.captureTimer = setInterval(() => {
-      this._flushSegment().catch((err) => {
-        console.error('[FriendService] segment timer error:', err);
-      });
-    }, this.SEGMENT_MS);
-  }
-
-  private _clearSegmentTimer(): void {
-    if (this.captureTimer) {
-      clearInterval(this.captureTimer);
-      this.captureTimer = null;
-    }
-  }
-
   /** Race a promise against a timeout */
   private async withTimeout<T>(
     promise: Promise<T>,
@@ -403,16 +408,20 @@ class FriendService {
     console.log(`[FriendService] _stopCapture: capturing=${this.capturing} sttConnection=${this.sttConnection ? 'exists' : 'null'}`);
     if (!this.capturing) return '';
 
-    this.capturing = false;
-
-    // Stop segment timer first
-    this._clearSegmentTimer();
-
-    // Stop cpal recording
+    // Stop audio recording first (no more audio → no more STT/VAD input)
     if (this.audioCapture) {
       await this.audioCapture.stopRecording().catch(() => {});
     }
 
+    this.capturing = false;
+
+    // Reset VAD state silently (don't fire onSpeechEnd — we're finalizing below)
+    if (this.vadInstance) {
+      try {
+        this.vadInstance.reset();
+      } catch { /* ignore */ }
+    }
+
     // Finalize STT connection
     const conn = this.sttConnection;
     this.sttConnection = null;
@@ -615,13 +624,28 @@ class FriendService {
             let verifyTimer: ReturnType<typeof setTimeout> | null = null;
 
             const verified = await new Promise<boolean>((resolve) => {
+              const feedVad = (c: Buffer) => {
+                if (this.vadInstance) {
+                  const float32 = new Float32Array(c.length / 2);
+                  for (let i = 0; i < float32.length; i++) {
+                    float32[i] = c.readInt16LE(i * 2) / 32768;
+                  }
+                  this.vadInstance.processAudio(float32).catch(() => {});
+                }
+              };
+
               const dataHandler = (chunk: Buffer) => {
                 dataArrived = true;
                 if (verifyTimer) { clearTimeout(verifyTimer); }
-                // Forward first chunk, then swap to the permanent handler
+                // Forward first chunk to both STT and VAD
                 onData(chunk);
+                feedVad(chunk);
+                // Swap to the permanent handler for subsequent chunks
                 proc.stdout?.removeListener('data', dataHandler);
-                proc.stdout?.on('data', onData);
+                proc.stdout?.on('data', (c: Buffer) => {
+                  onData(c);
+                  feedVad(c);
+                });
                 resolve(true);
               };
               proc.stdout?.on('data', dataHandler);

src/friend/voice/vad-service.ts

@@ -0,0 +1,290 @@
+/**
+ * Silero VAD service — real-time voice activity detection using onnxruntime-web WASM backend.
+ *
+ * Uses the Silero VAD legacy ONNX model (bundled with @ericedouard/vad-node-realtime).
+ * onnxruntime-web (WASM) is used instead of onnxruntime-node (native addon) because
+ * Bun does not support the native Node-API addon (crashes with segfault).
+ *
+ * Architecture:
+ *   - Audio frames (512 samples @ 16kHz = 32ms) are fed to the Silero model
+ *   - Speech probability is compared against configurable thresholds
+ *   - A state machine tracks speech segments with redemption grace period
+ *   - onSpeechEnd fires when sustained silence is detected
+ */
+
+import * as ort from 'onnxruntime-web';
+import { readFileSync, existsSync } from 'node:fs';
+import { resolve, dirname } from 'node:path';
+
+export interface VadCallbacks {
+  onSpeechStart?: () => void;
+  onSpeechEnd?: (audio: Float32Array) => void;
+  onVADMisfire?: () => void;
+  onFrameProcessed?: (prob: number, isSpeech: boolean) => void;
+}
+
+export interface VadOptions {
+  /** Threshold above which a frame is considered speech (0-1). Default: 0.5 */
+  positiveSpeechThreshold?: number;
+  /** Threshold below which a frame is considered silence (0-1). Default: 0.35 */
+  negativeSpeechThreshold?: number;
+  /** Consecutive silence frames before onSpeechEnd fires. Default: 8 (~256ms) */
+  redemptionFrames?: number;
+  /** Minimum speech frames to avoid misfire. Default: 3 (~96ms) */
+  minSpeechFrames?: number;
+  /** Frames of pre-speech audio to include in onSpeechEnd segment. Default: 10 */
+  preSpeechPadFrames?: number;
+  /** Sample rate of input audio (must be 16000). Default: 16000 */
+  sampleRate?: number;
+}
+
+export class SileroVad {
+  private session: ort.InferenceSession | null = null;
+  private stateH: ort.Tensor<onnxruntime.TensorType> | null = null;
+  private stateC: ort.Tensor<onnxruntime.TensorType> | null = null;
+  private sr: ort.Tensor<onnxruntime.TensorType> | null = null;
+  private initialized = false;
+  private active = false;
+
+  private readonly opts: Required<VadOptions>;
+  private readonly callbacks: Required<VadCallbacks>;
+  private readonly frameSize = 512; // Legacy Silero model: 512 samples @ 16kHz per frame
+
+  // Audio accumulation buffer
+  private buffer = new Float32Array(0);
+
+  // Frame processor state machine
+  private speaking = false;
+  private redemptionCounter = 0;
+  private speechFrameCount = 0;
+  private frameHistory: Array<{ frame: Float32Array; isSpeech: boolean }> = [];
+
+  constructor(callbacks: VadCallbacks, opts?: VadOptions) {
+    this.callbacks = {
+      onSpeechStart: callbacks.onSpeechStart ?? (() => {}),
+      onSpeechEnd: callbacks.onSpeechEnd ?? (() => {}),
+      onVADMisfire: callbacks.onVADMisfire ?? (() => {}),
+      onFrameProcessed: callbacks.onFrameProcessed ?? (() => {}),
+    };
+    this.opts = {
+      positiveSpeechThreshold: opts?.positiveSpeechThreshold ?? 0.5,
+      negativeSpeechThreshold: opts?.negativeSpeechThreshold ?? 0.35,
+      redemptionFrames: opts?.redemptionFrames ?? 8,
+      minSpeechFrames: opts?.minSpeechFrames ?? 3,
+      preSpeechPadFrames: opts?.preSpeechPadFrames ?? 10,
+      sampleRate: opts?.sampleRate ?? 16000,
+    };
+  }
+
+  /**
+   * Initialize the VAD: load ONNX model and configure onnxruntime-web WASM backend.
+   * Must be called once before start().
+   */
+  async init(): Promise<void> {
+    if (this.initialized) return;
+
+    // Point onnxruntime-web to the WASM binary files
+    const wasmDir = resolveWasmDir();
+    ort.env.wasm.wasmPaths = wasmDir + '/';
+
+    // Load Silero VAD legacy ONNX model
+    const modelPath = resolveModelPath();
+    const modelBuffer = readFileSync(modelPath);
+    this.session = await ort.InferenceSession.create(modelBuffer.buffer, {
+      executionProviders: ['wasm'],
+    });
+
+    // Initialize LSTM state tensors (legacy: h=[2,1,64], c=[2,1,64])
+    this.stateH = new ort.Tensor('float32', new Float32Array(2 * 64), [2, 1, 64]);
+    this.stateC = new ort.Tensor('float32', new Float32Array(2 * 64), [2, 1, 64]);
+    this.sr = new ort.Tensor('int64', [BigInt(this.opts.sampleRate)], [1]);
+
+    this.initialized = true;
+  }
+
+  /** Start VAD processing */
+  start(): void {
+    this.active = true;
+  }
+
+  /** Pause VAD processing, ending any active speech segment */
+  pause(): void {
+    this.active = false;
+    this.endSegment();
+  }
+
+  /**
+   * Feed raw PCM audio (Float32Array, values -1..1, 16kHz) to the VAD.
+   * Audio is buffered and processed in 512-sample frames.
+   */
+  async processAudio(audioData: Float32Array): Promise<void> {
+    if (!this.active || !this.initialized || !this.session) return;
+
+    // Append to internal buffer
+    const tmp = new Float32Array(this.buffer.length + audioData.length);
+    tmp.set(this.buffer);
+    tmp.set(audioData, this.buffer.length);
+    this.buffer = tmp;
+
+    // Process complete 512-sample frames
+    while (this.buffer.length >= this.frameSize) {
+      const frame = this.buffer.subarray(0, this.frameSize);
+      this.buffer = this.buffer.subarray(this.frameSize);
+      await this.processFrame(frame);
+    }
+  }
+
+  /** Flush any remaining audio and end active speech segment */
+  async flush(): Promise<void> {
+    if (this.buffer.length > 0) {
+      // Pad last partial frame with zeros
+      const padded = new Float32Array(this.frameSize);
+      padded.set(this.buffer);
+      this.buffer = new Float32Array(0);
+      await this.processFrame(padded);
+    }
+    this.endSegment();
+  }
+
+  /** Reset VAD state without destroying the session */
+  reset(): void {
+    this.buffer = new Float32Array(0);
+    this.frameHistory = [];
+    this.speaking = false;
+    this.redemptionCounter = 0;
+    this.speechFrameCount = 0;
+    this.stateH = new ort.Tensor('float32', new Float32Array(2 * 64), [2, 1, 64]);
+    this.stateC = new ort.Tensor('float32', new Float32Array(2 * 64), [2, 1, 64]);
+  }
+
+  /** Clean up resources */
+  destroy(): void {
+    this.active = false;
+    this.reset();
+    if (this.session) {
+      try { (this.session as any).dispose?.(); } catch { /* ignore */ }
+      this.session = null;
+    }
+    this.initialized = false;
+  }
+
+  // ── Private: frame processing ──────────────────────────────────────────
+
+  private async processFrame(frame: Float32Array): Promise<void> {
+    if (!this.session || !this.stateH || !this.stateC || !this.sr) return;
+
+    try {
+      const result = await this.session.run({
+        input: new ort.Tensor('float32', frame, [1, this.frameSize]),
+        sr: this.sr,
+        h: this.stateH,
+        c: this.stateC,
+      });
+
+      // Update LSTM state for next frame
+      this.stateH = result.hn as ort.Tensor<onnxruntime.TensorType>;
+      this.stateC = result.cn as ort.Tensor<onnxruntime.TensorType>;
+
+      const prob = (result.output as ort.Tensor<onnxruntime.TensorType>).data[0] as number;
+      const isSpeech = prob >= this.opts.positiveSpeechThreshold;
+      const isSilence = prob < this.opts.negativeSpeechThreshold;
+
+      this.callbacks.onFrameProcessed(prob, isSpeech);
+
+      // ── State machine ──────────────────────────────────────────────
+      if (this.speaking) {
+        // Currently in a speech segment
+        if (isSilence) {
+          this.redemptionCounter++;
+          if (this.redemptionCounter >= this.opts.redemptionFrames) {
+            this.endSpeech();
+          }
+        } else {
+          this.redemptionCounter = 0;
+        }
+        this.speechFrameCount++;
+        this.frameHistory.push({ frame: frame.slice(), isSpeech });
+      } else if (isSpeech) {
+        // Transition: silence → speech
+        this.speaking = true;
+        this.redemptionCounter = 0;
+        this.speechFrameCount = 1;
+
+        // Include pre-padding frames (audio just before speech start)
+        this.frameHistory.push({ frame: frame.slice(), isSpeech: true });
+        this.callbacks.onSpeechStart();
+      }
+    } catch (err) {
+      console.error('[SileroVad] frame inference error:', err);
+    }
+  }
+
+  private endSpeech(): void {
+    if (this.speechFrameCount < this.opts.minSpeechFrames) {
+      this.callbacks.onVADMisfire();
+    } else {
+      // Build audio segment with pre-padding for context
+      const total = this.frameHistory.length;
+      const prePad = Math.min(this.opts.preSpeechPadFrames, total);
+      const segFrames = this.frameHistory.slice(total - prePad - this.speechFrameCount, total);
+      let totalSamples = 0;
+      for (const f of segFrames) totalSamples += f.frame.length;
+      const segment = new Float32Array(totalSamples);
+      let offset = 0;
+      for (const f of segFrames) {
+        segment.set(f.frame, offset);
+        offset += f.frame.length;
+      }
+      this.callbacks.onSpeechEnd(segment);
+    }
+
+    // Reset speech state
+    this.speaking = false;
+    this.redemptionCounter = 0;
+    this.speechFrameCount = 0;
+    this.frameHistory = [];
+  }
+
+  private endSegment(): void {
+    if (this.speaking) {
+      this.endSpeech();
+    }
+  }
+}
+
+// ── Module-level helpers ─────────────────────────────────────────────────
+
+let cachedWasmDir: string | null = null;
+let cachedModelPath: string | null = null;
+
+function resolveWasmDir(): string {
+  if (cachedWasmDir) return cachedWasmDir;
+
+  // Resolve onnxruntime-web's dist directory where WASM files live
+  const pkgPath = require.resolve('onnxruntime-web/package.json');
+  cachedWasmDir = resolve(dirname(pkgPath), 'dist');
+
+  if (!existsSync(cachedWasmDir)) {
+    throw new Error(
+      `onnxruntime-web WASM directory not found at ${cachedWasmDir}. ` +
+        'Ensure onnxruntime-web is installed (it is a peer dependency).',
+    );
+  }
+  return cachedWasmDir;
+}
+
+function resolveModelPath(): string {
+  if (cachedModelPath) return cachedModelPath;
+
+  // The Silero ONNX model is bundled with @ericedouard/vad-node-realtime
+  const vadPkgPath = require.resolve('@ericedouard/vad-node-realtime/package.json');
+  cachedModelPath = resolve(dirname(vadPkgPath), 'silero_vad_legacy.onnx');
+
+  if (!existsSync(cachedModelPath)) {
+    throw new Error(
+      `Silero VAD model not found at ${cachedModelPath}. ` +
+        'Ensure @ericedouard/vad-node-realtime is installed.',
+    );
+  }
+  return cachedModelPath;
+}