Prepare counted Hugging Face model downloads

README.md

@@ -22,7 +22,10 @@ Audio and transcripts are never sent to an inference server. The most recent
 saved.
 
 The model URLs and checksums are pinned in [`model-config.json`](model-config.json).
-The first load downloads Tiny and then PnC. Microphone access requires a secure
+For the public release, the browser downloads Tiny's runtime configuration and
+ONNX files, then PnC's files, directly from their Hugging Face model repos. The
+private preview serves verified local copies while those repos remain private.
+Microphone access requires a secure
 origin and the browser's permission prompt, shown after the app's introductory
 mic dialog. Browser support remains a preview while physical-device and mobile
 checks are completed.

model-config.json

@@ -2,11 +2,8 @@
   "schema": "ekko-browser-model-v1",
   "name": "Ekko v1 Tiny int8",
   "repository": "RyeAI/ekko-v1-tiny",
-  "revision": "f1070717d45491f941a1486affe53fef5f7104ad",
+  "revision": "e37e698805c538e447c25062ab13fa862462ee9e",
   "runtime": "sherpa-onnx 1.13.6 WebAssembly",
-  "sampleRate": 16000,
-  "featureDim": 80,
-  "modelType": "nemo_transducer",
   "pnc": {
     "name": "Ekko PnC v2 int8",
     "repository": "RyeAI/ekko-pnc",
@@ -35,6 +32,12 @@
     ]
   },
   "files": [
+    {
+      "path": "runtime-config.json",
+      "url": "model-files/tiny/runtime-config.json?sha256=197ec64ea58975cb70c8bac645593bbaaa7ac7b261999b35b7d2986a26a41da5",
+      "sha256": "197ec64ea58975cb70c8bac645593bbaaa7ac7b261999b35b7d2986a26a41da5",
+      "bytes": 121
+    },
     {
       "path": "nemo-transducer-encoder.onnx",
       "url": "model-files/tiny/nemo-transducer-encoder.onnx?sha256=e7ee366cb92cbf5b14beee01f04245293185ec3b271bb12acd7715e94d1f4cf2",

model-files/tiny/runtime-config.json

@@ -0,0 +1,6 @@
+{
+  "schema": "ekko-tiny-onnx-runtime-v1",
+  "sampleRate": 16000,
+  "featureDim": 80,
+  "modelType": "nemo_transducer"
+}

validation.json

@@ -13,7 +13,7 @@
     "bundle_bytes_approximate": 13600000
   },
   "model": {
-    "revision": "f1070717d45491f941a1486affe53fef5f7104ad",
+    "revision": "e37e698805c538e447c25062ab13fa862462ee9e",
     "download_bytes": 135307048,
     "local_verified_cache_server": true,
     "reason": "The private v5 artifact revision loaded through the checksum-verified local cache server."
@@ -178,7 +178,7 @@
     },
     "v5_browser_wasm_smoke": {
       "status": "passed-local-private-preview",
-      "model_revision": "f1070717d45491f941a1486affe53fef5f7104ad",
+      "model_revision": "e37e698805c538e447c25062ab13fa862462ee9e",
       "int8_bundle_bytes": 135307058,
       "current_matrix_mean_wer": 12.945774,
       "one_word_fixture_seconds": 0.47,

worker.js

@@ -179,12 +179,12 @@ async function fetchVerified(file, completedBytes, totalBytes) {
   return bytes;
 }
 
-function createRecognizer(config) {
+function createRecognizer(runtimeConfig) {
   recognizer = new OfflineRecognizer(
     {
       featConfig: {
-        sampleRate: config.sampleRate,
-        featureDim: config.featureDim,
+        sampleRate: runtimeConfig.sampleRate,
+        featureDim: runtimeConfig.featureDim,
       },
       modelConfig: {
         transducer: {
@@ -193,7 +193,7 @@ function createRecognizer(config) {
           joiner: "./nemo-transducer-joiner.onnx",
         },
         tokens: "./tokens.txt",
-        modelType: config.modelType,
+        modelType: runtimeConfig.modelType,
         provider: "cpu",
         numThreads: 1,
         debug: 0,
@@ -233,12 +233,25 @@ async function initialize(configUrl) {
 
   const totalBytes = config.files.reduce((sum, file) => sum + file.bytes, 0);
   let completedBytes = 0;
+  let runtimeConfig = null;
   for (const file of config.files) {
     const bytes = await fetchVerified(file, completedBytes, totalBytes);
-    Module.FS.writeFile(file.path, bytes);
+    if (file.path === "runtime-config.json") {
+      runtimeConfig = JSON.parse(new TextDecoder().decode(bytes));
+    } else {
+      Module.FS.writeFile(file.path, bytes);
+    }
     completedBytes += bytes.byteLength;
   }
-  createRecognizer(config);
+  if (
+    runtimeConfig?.schema !== "ekko-tiny-onnx-runtime-v1" ||
+    runtimeConfig.sampleRate !== SAMPLE_RATE ||
+    runtimeConfig.featureDim !== 80 ||
+    runtimeConfig.modelType !== "nemo_transducer"
+  ) {
+    throw new Error("Ukendt Tiny runtimekonfiguration");
+  }
+  createRecognizer(runtimeConfig);
   self.postMessage({
     type: "ready",
     model: config.name,