@tryhamster/gerbil 1.12.0 → 1.13.0

package/README.md

@@ -349,6 +349,30 @@ const engine = await WebGPUEngine.create({ repo: "nineninesix/kani-tts-2-en" });
 const { pcm, sampleRate } = await engine.speak("Hello, I'm Gerbil!"); // sampleRate === 22050
 ```
 
+**Streaming TTS** — native **Gepard-1.0** (Apache-2.0). Gepard is a stock Qwen3.5
+codec-LM with 32 parallel FSQ classifier heads that emit one whole audio frame per
+forward pass, so speech is frame synchronous: audio streams out chunk by chunk while
+the model is still generating. Pass `onChunk` to receive PCM as it is produced.
+
+```typescript
+const engine = await WebGPUEngine.create({ repo: "nineninesix/gepard-1.0" });
+const { pcm, sampleRate } = await engine.speak("Hello from Gepard.", {
+  temperature: 0.4,
+  onChunk: ({ pcm, frameIndex, isFinal }) => {
+    // play or buffer each ~1.5s chunk as soon as it is decoded
+  },
+}); // sampleRate === 22050
+```
+
+Run the end to end example and write a .wav:
+
+```bash
+node scripts/engine/example-gepard-tts.mjs "The quick brown fox jumps over the lazy dog."
+```
+
+Voice cloning (the reference speaker compressor) is present in the checkpoint but not
+wired yet, so Gepard ships a single default voice for now.
+
 **Speech-to-text** — native **Moonshine** (raw-waveform encoder/decoder, no FFT/log-mel)
 via the dedicated `MoonshineSTT` class.
 
@@ -569,6 +593,7 @@ via `WebGPUEngine.create({ repo })`.
 | Model | Type | Repo | Notes |
 |-------|------|------|-------|
 | **Kani-TTS-2** | TTS | `nineninesix/kani-tts-2-en` | `engine.speak()` → 22.05 kHz PCM |
+| **Gepard-1.0** | TTS | `nineninesix/gepard-1.0` | Streaming `engine.speak({ onChunk })` → 22.05 kHz PCM, Apache-2.0 |
 | **Moonshine** | STT | `UsefulSensors/moonshine-base` | `MoonshineSTT.transcribe()`, raw-waveform |
 
 ### Quantization & dtype

package/dist/{architectures-BH_z6k9d.mjs → architectures-DHwj9AQD.mjs}

@@ -445,6 +445,8 @@ function ncWeight(name) {
 }
 function generateNanoCodecDecoderGraph(opts) {
 	const T = opts.numFrames;
+	const upSampleRates = opts.geometry?.upSampleRates ?? NANOCODEC_UP_SAMPLE_RATES;
+	const inputDim = opts.geometry?.inputDim ?? NANOCODEC_INPUT_DIM;
 	const tensors = {};
 	const nodes = [];
 	const executionOrder = [];
@@ -455,31 +457,39 @@ function generateNanoCodecDecoderGraph(opts) {
 		nodes.push(node);
 		executionOrder.push(node.id);
 	};
-	addTensor({
-		name: "audio_codes",
-		shape: [FSQ_NUM_GROUPS, T],
-		dtype: "u32",
-		storage: "activation"
-	});
-	addTensor({
+	if (opts.latentInput) addTensor({
 		name: "nc_latent",
-		shape: [NANOCODEC_INPUT_DIM, T],
+		shape: [inputDim, T],
 		dtype: "f32",
 		storage: "activation"
 	});
-	addNode({
-		id: "nc_fsq",
-		opType: "FSQDequant",
-		inputs: ["audio_codes"],
-		outputs: ["nc_latent"],
-		attributes: {
-			groups: FSQ_NUM_GROUPS,
-			dims: FSQ_DIMS_PER_GROUP,
-			T,
-			levels: [...FSQ_LEVELS],
-			base: [...FSQ_BASE]
-		}
-	});
+	else {
+		addTensor({
+			name: "audio_codes",
+			shape: [FSQ_NUM_GROUPS, T],
+			dtype: "u32",
+			storage: "activation"
+		});
+		addTensor({
+			name: "nc_latent",
+			shape: [inputDim, T],
+			dtype: "f32",
+			storage: "activation"
+		});
+		addNode({
+			id: "nc_fsq",
+			opType: "FSQDequant",
+			inputs: ["audio_codes"],
+			outputs: ["nc_latent"],
+			attributes: {
+				groups: FSQ_NUM_GROUPS,
+				dims: FSQ_DIMS_PER_GROUP,
+				T,
+				levels: [...FSQ_LEVELS],
+				base: [...FSQ_BASE]
+			}
+		});
+	}
 	/** Emit a causal Conv1dFull (padding = (K-1)*dilation, Lout = L). */
 	function causalConv(id, input, weightKey, biasKey, Cin, Cout, K, L, dilation, output) {
 		addTensor({
@@ -619,11 +629,11 @@ function generateNanoCodecDecoderGraph(opts) {
 		});
 	}
 	let cur = "nc_pre_conv_out";
-	causalConv("nc_pre_conv", "nc_latent", ncWeight("pre_conv.weight"), ncWeight("pre_conv.bias"), NANOCODEC_INPUT_DIM, NANOCODEC_BASE_CHANNELS, 7, T, 1, cur);
+	causalConv("nc_pre_conv", "nc_latent", ncWeight("pre_conv.weight"), ncWeight("pre_conv.bias"), inputDim, NANOCODEC_BASE_CHANNELS, 7, T, 1, cur);
 	let channels = NANOCODEC_BASE_CHANNELS;
 	let length = T;
-	for (let stage = 0; stage < NANOCODEC_UP_SAMPLE_RATES.length; stage++) {
-		const rate = NANOCODEC_UP_SAMPLE_RATES[stage];
+	for (let stage = 0; stage < upSampleRates.length; stage++) {
+		const rate = upSampleRates[stage];
 		const outCh = channels / 2;
 		const upK = 2 * rate;
 		const hsOut = `nc_s${stage}_act`;
@@ -681,7 +691,7 @@ function generateNanoCodecDecoderGraph(opts) {
 		architecture: "NanoCodecDecoder",
 		config: {
 			hidden_size: NANOCODEC_BASE_CHANNELS,
-			num_layers: NANOCODEC_UP_SAMPLE_RATES.length,
+			num_layers: upSampleRates.length,
 			num_heads: 0,
 			num_kv_heads: 0,
 			head_dim: 0,
@@ -704,7 +714,7 @@ function generateNanoCodecDecoderGraph(opts) {
 		tensors,
 		nodes,
 		executionOrder,
-		inputs: ["audio_codes"],
+		inputs: [opts.latentInput ? "nc_latent" : "audio_codes"],
 		outputs: [postConv]
 	};
 }
@@ -6769,4 +6779,4 @@ function isArchitectureSupported(name) {
 
 //#endregion
 export { kaniSinTensor as A, computeKaniPositions as C, kaniAttentionLayerIndices as D, generateNanoCodecDecoderGraph as E, CANONICAL_KEYS as F, DTYPE_BYTES as I, GEMMA4_VIS_KEYS as L, parseKaniConfig as M, DEFAULT_GROUP_SIZE as N, kaniCosTensor as O, quantizeInt4 as P, createDefaultHFKeyMapper as R, buildKaniLayerCosSin as S, generateKaniTtsGraph as T, moonshineEncoderFrames as _, PARLER_DAC_LATENT_DIM as a, KANI_START_OF_HUMAN as b, PARLER_SAMPLE_RATE as c, generateParlerEncoderGraph as d, parseParlerConfig as f, generateMoonshineEncoderGraph as g, generateMoonshineDecoderGraph as h, PARLER_DAC_DECODER_DIM as i, nanoCodecWeightMap as j, kaniLayerAlpha as k, buildT5RelativeBias as l, MOONSHINE_REMAINING_WORK as m, isArchitectureSupported as n, PARLER_DECODER_RATES as o, revertDelayPattern as p, PARLER_BOS_TOKEN_ID as r, PARLER_EOS_TOKEN_ID as s, generateGraph as t, generateParlerDecoderGraph as u, parseMoonshineConfig as v, foldNanoCodecWeightNorm as w, audioTokensToCodes as x, KANI_END_OF_HUMAN as y };
-//# sourceMappingURL=architectures-BH_z6k9d.mjs.map
+//# sourceMappingURL=architectures-DHwj9AQD.mjs.map