@trillboards/edge-sdk 0.2.49 → 0.2.52

package/dist/wavEncode.d.ts

@@ -0,0 +1,29 @@
+/**
+ * wavEncode — Float32 PCM → a 16-bit mono RIFF/WAVE buffer.
+ *
+ * This exists because litert-lm's `input_audio` takes a CONTAINER (wav/flac/mp3
+ * base64), not raw samples, while everything upstream of it in this SDK —
+ * `AudioProvider.getAudioChunk`, `RTSPAudioProvider`'s ring, the onset
+ * detector — speaks Float32. One conversion, in one place, so the header is
+ * written correctly once.
+ *
+ * MEASURED CONSTRAINTS (litert-lm 0.16.0, `gemma-4-E2B-it.litertlm`):
+ * mono, 16 kHz, ≤30 s. Those are what `GEMMA_AUDIO_*` below encode. Speech at
+ * these settings transcribes verbatim (30/30 trials); acoustic events do not
+ * classify at all, which is why nothing in this SDK asks the model for one —
+ * see `AudioOnsetDetector`.
+ */
+/** Sample rate litert-lm's audio encoder expects. */
+export declare const GEMMA_AUDIO_SAMPLE_RATE = 16000;
+/** Longest clip litert-lm accepts, in seconds. */
+export declare const GEMMA_AUDIO_MAX_SECONDS = 30;
+/**
+ * Encode mono Float32 samples (nominally -1..1) as a 16-bit PCM WAV.
+ * Out-of-range samples are CLAMPED, not wrapped: a wrapped sample turns a loud
+ * transient into full-scale noise of the opposite sign, which is exactly the
+ * content an acoustic path is most likely to be handed.
+ */
+export declare function encodeWav16(samples: Float32Array, sampleRate: number): Buffer;
+/** The most recent `seconds` of a buffer, for the model's clip-length cap. */
+export declare function takeMostRecentSeconds(samples: Float32Array, sampleRate: number, seconds?: number): Float32Array;
+//# sourceMappingURL=wavEncode.d.ts.map

package/dist/wavEncode.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"wavEncode.d.ts","sourceRoot":"","sources":["../src/wavEncode.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,qDAAqD;AACrD,eAAO,MAAM,uBAAuB,QAAS,CAAC;AAC9C,kDAAkD;AAClD,eAAO,MAAM,uBAAuB,KAAK,CAAC;AAE1C;;;;;GAKG;AACH,wBAAgB,WAAW,CAAC,OAAO,EAAE,YAAY,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAyB7E;AAED,8EAA8E;AAC9E,wBAAgB,qBAAqB,CACnC,OAAO,EAAE,YAAY,EACrB,UAAU,EAAE,MAAM,EAClB,OAAO,GAAE,MAAgC,GACxC,YAAY,CAId"}

package/dist/wavEncode.js

@@ -0,0 +1,63 @@
+"use strict";
+/**
+ * wavEncode — Float32 PCM → a 16-bit mono RIFF/WAVE buffer.
+ *
+ * This exists because litert-lm's `input_audio` takes a CONTAINER (wav/flac/mp3
+ * base64), not raw samples, while everything upstream of it in this SDK —
+ * `AudioProvider.getAudioChunk`, `RTSPAudioProvider`'s ring, the onset
+ * detector — speaks Float32. One conversion, in one place, so the header is
+ * written correctly once.
+ *
+ * MEASURED CONSTRAINTS (litert-lm 0.16.0, `gemma-4-E2B-it.litertlm`):
+ * mono, 16 kHz, ≤30 s. Those are what `GEMMA_AUDIO_*` below encode. Speech at
+ * these settings transcribes verbatim (30/30 trials); acoustic events do not
+ * classify at all, which is why nothing in this SDK asks the model for one —
+ * see `AudioOnsetDetector`.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.GEMMA_AUDIO_MAX_SECONDS = exports.GEMMA_AUDIO_SAMPLE_RATE = void 0;
+exports.encodeWav16 = encodeWav16;
+exports.takeMostRecentSeconds = takeMostRecentSeconds;
+/** Sample rate litert-lm's audio encoder expects. */
+exports.GEMMA_AUDIO_SAMPLE_RATE = 16000;
+/** Longest clip litert-lm accepts, in seconds. */
+exports.GEMMA_AUDIO_MAX_SECONDS = 30;
+/**
+ * Encode mono Float32 samples (nominally -1..1) as a 16-bit PCM WAV.
+ * Out-of-range samples are CLAMPED, not wrapped: a wrapped sample turns a loud
+ * transient into full-scale noise of the opposite sign, which is exactly the
+ * content an acoustic path is most likely to be handed.
+ */
+function encodeWav16(samples, sampleRate) {
+    if (!(sampleRate > 0))
+        throw new Error('encodeWav16 requires a positive sampleRate');
+    const dataBytes = samples.length * 2;
+    const buf = Buffer.alloc(44 + dataBytes);
+    buf.write('RIFF', 0, 'ascii');
+    buf.writeUInt32LE(36 + dataBytes, 4);
+    buf.write('WAVE', 8, 'ascii');
+    buf.write('fmt ', 12, 'ascii');
+    buf.writeUInt32LE(16, 16); // PCM fmt chunk size
+    buf.writeUInt16LE(1, 20); // audioFormat = PCM
+    buf.writeUInt16LE(1, 22); // channels = mono
+    buf.writeUInt32LE(sampleRate, 24);
+    buf.writeUInt32LE(sampleRate * 2, 28); // byte rate (mono, 2 bytes/sample)
+    buf.writeUInt16LE(2, 32); // block align
+    buf.writeUInt16LE(16, 34); // bits per sample
+    buf.write('data', 36, 'ascii');
+    buf.writeUInt32LE(dataBytes, 40);
+    for (let i = 0; i < samples.length; i++) {
+        const clamped = Math.max(-1, Math.min(1, samples[i]));
+        // Asymmetric scaling: int16 spans -32768..32767, so 1.0 must map to 32767.
+        buf.writeInt16LE(Math.round(clamped * (clamped < 0 ? 32768 : 32767)), 44 + i * 2);
+    }
+    return buf;
+}
+/** The most recent `seconds` of a buffer, for the model's clip-length cap. */
+function takeMostRecentSeconds(samples, sampleRate, seconds = exports.GEMMA_AUDIO_MAX_SECONDS) {
+    const max = Math.max(1, Math.round(sampleRate * seconds));
+    if (samples.length <= max)
+        return samples;
+    return samples.slice(samples.length - max);
+}
+//# sourceMappingURL=wavEncode.js.map

package/dist/wavEncode.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"wavEncode.js","sourceRoot":"","sources":["../src/wavEncode.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;GAcG;;;AAaH,kCAyBC;AAGD,sDAQC;AA/CD,qDAAqD;AACxC,QAAA,uBAAuB,GAAG,KAAM,CAAC;AAC9C,kDAAkD;AACrC,QAAA,uBAAuB,GAAG,EAAE,CAAC;AAE1C;;;;;GAKG;AACH,SAAgB,WAAW,CAAC,OAAqB,EAAE,UAAkB;IACnE,IAAI,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,4CAA4C,CAAC,CAAC;IACrF,MAAM,SAAS,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC;IACrC,MAAM,GAAG,GAAG,MAAM,CAAC,KAAK,CAAC,EAAE,GAAG,SAAS,CAAC,CAAC;IAEzC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,OAAO,CAAC,CAAC;IAC9B,GAAG,CAAC,aAAa,CAAC,EAAE,GAAG,SAAS,EAAE,CAAC,CAAC,CAAC;IACrC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,OAAO,CAAC,CAAC;IAC9B,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,EAAE,EAAE,OAAO,CAAC,CAAC;IAC/B,GAAG,CAAC,aAAa,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,qBAAqB;IAChD,GAAG,CAAC,aAAa,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,oBAAoB;IAC9C,GAAG,CAAC,aAAa,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,kBAAkB;IAC5C,GAAG,CAAC,aAAa,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC;IAClC,GAAG,CAAC,aAAa,CAAC,UAAU,GAAG,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,mCAAmC;IAC1E,GAAG,CAAC,aAAa,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,cAAc;IACxC,GAAG,CAAC,aAAa,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,kBAAkB;IAC7C,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,EAAE,EAAE,OAAO,CAAC,CAAC;IAC/B,GAAG,CAAC,aAAa,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;IAEjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACxC,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,2EAA2E;QAC3E,GAAG,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,GAAG,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;IACpF,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,8EAA8E;AAC9E,SAAgB,qBAAqB,CACnC,OAAqB,EACrB,UAAkB,EAClB,UAAkB,+BAAuB;IAEzC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,UAAU,GAAG,OAAO,CAAC,CAAC,CAAC;IAC1D,IAAI,OAAO,CAAC,MAAM,IAAI,GAAG;QAAE,OAAO,OAAO,CAAC;IAC1C,OAAO,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC;AAC7C,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@trillboards/edge-sdk",
-  "version": "0.2.49",
+  "version": "0.2.52",
   "description": "Composable cross-platform Edge AI SDK for audience sensing on DOOH/CTV devices",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",