@datalayer/agent-runtimes 1.3.62 → 1.3.64

package/scripts/voice/serve_store.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+# Copyright (c) 2025-2026 Datalayer, Inc.
+# Distributed under the terms of the Modified BSD License.
+
+"""Serve the pinned store to a page on this machine, as Datalayer's origin serves it (VOICE.md VO-49).
+
+    python scripts/voice/serve_store.py --store ~/.cache/speech-store [--port 8770]
+
+The browser's models at ``/<model id>/<path>`` and onnxruntime-web's own
+files at ``/onnxruntime-web/<version>/``, from the onnxruntime-web the page
+bundles (``--node-modules`` when it is not this checkout's), with the CORS
+headers a page on another port needs. For a local run only: in production
+the same layout is on Datalayer's storage.
+"""
+
+from __future__ import annotations
+
+import argparse
+import functools
+import json
+from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+class StoreHandler(SimpleHTTPRequestHandler):
+    runtime: Path = Path()
+    version: str = ""
+
+    def translate_path(self, path: str) -> str:
+        prefix = f"/onnxruntime-web/{self.version}/"
+        if path.startswith(prefix):
+            return str(self.runtime / path[len(prefix) :].split("?")[0])
+        return super().translate_path(path)
+
+    def end_headers(self) -> None:
+        self.send_header("Access-Control-Allow-Origin", "*")
+        self.send_header("Cross-Origin-Resource-Policy", "cross-origin")
+        super().end_headers()
+
+    def log_message(self, format: str, *args: object) -> None:  # noqa: A002
+        pass
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    parser.add_argument("--store", type=Path, required=True)
+    parser.add_argument("--port", type=int, default=8770)
+    parser.add_argument(
+        "--node-modules", type=Path, default=ROOT.parent.parent / "node_modules"
+    )
+    args = parser.parse_args()
+    runtime = args.node_modules / "onnxruntime-web"
+    if not (runtime / "package.json").is_file():
+        raise SystemExit(
+            f"No onnxruntime-web in {args.node_modules}: install the voice packages first."
+        )
+    StoreHandler.runtime = runtime / "dist"
+    StoreHandler.version = json.loads((runtime / "package.json").read_text())["version"]
+    handler = functools.partial(StoreHandler, directory=str(args.store))
+    print(
+        f"http://127.0.0.1:{args.port}/ serves {args.store} and onnxruntime-web {StoreHandler.version}"
+    )
+    ThreadingHTTPServer(("127.0.0.1", args.port), handler).serve_forever()
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/scripts/voice/transcribe.mjs

@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2025-2026 Datalayer, Inc.
+ * Distributed under the terms of the Modified BSD License.
+ */
+
+/**
+ * Transcribe WAV files with a speech-to-text model of the catalogue, as the
+ * browser runs it: transformers.js, the same ONNX files, read from the
+ * pinned store (VOICE.md VO-05, VO-51). Prints one JSON line.
+ *
+ *   node scripts/voice/transcribe.mjs <store> <model id> <language> <file.wav>...
+ *
+ * In Node the runtime is onnxruntime-node on the CPU, not the browser's
+ * WASM: the words are the model's, the timings are not the browser's.
+ */
+
+import { readFileSync } from 'node:fs';
+import { env, pipeline } from '@huggingface/transformers';
+
+const [, , store, model, language, ...files] = process.argv;
+env.allowRemoteModels = false;
+env.localModelPath = store.endsWith('/') ? store : `${store}/`;
+
+/** A 16-bit PCM mono WAV, as samples between -1 and 1. */
+function readWav(path) {
+  const bytes = readFileSync(path);
+  const at = bytes.indexOf('data') + 8;
+  const count = (bytes.length - at) / 2;
+  const samples = new Float32Array(count);
+  for (let index = 0; index < count; index += 1) {
+    samples[index] = bytes.readInt16LE(at + 2 * index) / 32768;
+  }
+  return samples;
+}
+
+const loading = performance.now();
+const transcribe = await pipeline('automatic-speech-recognition', model, {
+  dtype: 'q8',
+  device: 'cpu',
+});
+const loadMs = Math.round(performance.now() - loading);
+// Whisper is told the language; Moonshine hears English only.
+const options = model.startsWith('whisper')
+  ? { language: language === 'fr' ? 'french' : 'english', task: 'transcribe' }
+  : {};
+const results = [];
+for (const file of files) {
+  const audio = readWav(file);
+  const started = performance.now();
+  const output = await transcribe(audio, options);
+  results.push({
+    file: file.split('/').pop(),
+    text: output.text.trim(),
+    ms: Math.round(performance.now() - started),
+    audio_s: Math.round((audio.length / 16000) * 100) / 100,
+  });
+}
+console.log(JSON.stringify({ model, language, load_ms: loadMs, results }));