@tryhamster/gerbil 1.11.0 → 1.11.2

package/dist/{defaults-C_bJK9zs.mjs → defaults-DfGx4d1m.mjs}

@@ -47,6 +47,26 @@ const OUTETTS_PRESET_VOICES = [
 	"gerbil-voice-5",
 	"gerbil-voice-6"
 ];
+/**
+* True when `repo`/`architecture` identify a text-to-speech checkpoint that must
+* run through a dedicated native TTS engine (KaniTTS, OuteTTS, or Parler-TTS)
+* rather than the single-graph base text loader. Mirrors the routing used by
+* `WebGPUEngine.speak()` / `Gerbil.ensureNativeTTSEngine`:
+*   - Parler  → encoder-decoder (`ParlerTTSForConditionalGeneration`)
+*   - OuteTTS → Qwen3-0.6B codec-LM
+*   - Kani    → conv-backbone codec-LM (`KaniTTS2ForCausalLM`, or the Apache
+*               default which is a plain `Lfm2ForCausalLM` — hence the repo-name
+*               gate; the architecture string alone does not identify it).
+* At `create()` time only the repo is known (the architecture is discovered
+* during load), so the repo-name signals below cover all three DEFAULT_MODELS
+* and conventionally-named TTS repos; the architecture arg is honored when a
+* caller already has it.
+*/
+function isTTSRepo(repo, architecture) {
+	const r = repo ?? "";
+	const a = architecture ?? "";
+	return r === DEFAULT_MODELS.tts || r === DEFAULT_MODELS.ttsOute || r === DEFAULT_MODELS.ttsParler || /parler/i.test(r) || /parler/i.test(a) || /outetts/i.test(r) || /oute/i.test(a) || a === "KaniTTS2ForCausalLM" || /kani-tts/i.test(r) || /\btts\b/i.test(r);
+}
 /** Resolve the model repo for a set of options, falling back to the defaults. */
 function resolveDefaultRepo(opts) {
 	if (opts.repo) return opts.repo;
@@ -56,5 +76,5 @@ function resolveDefaultRepo(opts) {
 }
 
 //#endregion
-export { resolveDefaultRepo as i, OUTETTS_ASSETS as n, OUTETTS_PRESET_VOICES as r, DEFAULT_MODELS as t };
-//# sourceMappingURL=defaults-C_bJK9zs.mjs.map
+export { resolveDefaultRepo as a, isTTSRepo as i, OUTETTS_ASSETS as n, OUTETTS_PRESET_VOICES as r, DEFAULT_MODELS as t };
+//# sourceMappingURL=defaults-DfGx4d1m.mjs.map

package/dist/{defaults-C_bJK9zs.mjs.map → defaults-DfGx4d1m.mjs.map}

@@ -1 +1 @@
-{"version":3,"file":"defaults-C_bJK9zs.mjs","names":[],"sources":["../src/gpu/defaults.ts"],"sourcesContent":["/**\n * Default model per capability. Kept in its own tiny module (no heavy imports)\n * so the React hooks can resolve defaults without statically pulling in the GPU\n * engine — they import the engine dynamically to stay light.\n */\nexport const DEFAULT_MODELS = {\n  /** Text generation (also the vision-capable checkpoint). */\n  text: \"mlx-community/Qwen3.5-0.8B-4bit\",\n  /** Image understanding — same checkpoint, vision tower built on demand. */\n  vision: \"mlx-community/Qwen3.5-0.8B-4bit\",\n  /** Text embeddings. */\n  embedding: \"mlx-community/embeddinggemma-300m-4bit\",\n  /** Text-to-speech (Apache-2.0; LFM2-350M backbone + NanoCodec). Small/fast default. */\n  tts: \"nineninesix/kani-tts-450m-0.2-ft\",\n  /**\n   * Selectable many-voices TTS (OuteTTS-1.0-0.6B: Qwen3-0.6B codec-LM + DAC.speech).\n   * Pass this as the `repo`/`model` to engine.speak()/useTTS to route to OuteTTS.\n   */\n  ttsOute: \"OuteAI/OuteTTS-1.0-0.6B\",\n  /**\n   * Selectable describe-your-voice TTS (Parler-TTS-mini-v1: Flan-T5 encoder +\n   * cross-attending decoder LM + dac_44khz). Pass this as the `repo`/`model` to\n   * engine.speak()/useTTS with a `describeVoice` natural-language description.\n   */\n  ttsParler: \"parler-tts/parler-tts-mini-v1\",\n  /** Speech-to-text. */\n  stt: \"UsefulSensors/moonshine-base\",\n} as const;\n\n/**\n * OuteTTS auxiliary assets. The DAC.speech codec ships only `.pth` (non-safetensors),\n * so the folded weights (prep-dac-speech.py output) and the preset speaker JSONs are\n * hosted on the Cloudflare R2 model mirror (same public host the engine reads via\n * GERBIL_MODEL_BASE — range + CORS, free egress, no redirect). Override via the OuteTTS\n * options when self-hosting.\n */\nconst OUTETTS_ASSET_BASE =\n  \"https://pub-0522ecec25fc412bb35a523f3e3f63c6.r2.dev/tryhamster/gerbil-outetts-assets\";\nexport const OUTETTS_ASSETS = {\n  /** Folded DAC.speech.v1.0 encoder+decoder weights (prep-dac-speech.py output). */\n  codecWeightsURL: `${OUTETTS_ASSET_BASE}/dac-speech-folded.safetensors`,\n  /** Base URL for the preset speaker JSONs (`{name}.json`). */\n  speakersBaseURL: `${OUTETTS_ASSET_BASE}/speakers`,\n  /** The default preset voice. */\n  defaultVoice: \"en-female-1-neutral\",\n} as const;\n\n/**\n * Bundled OuteTTS preset voices.\n *\n * Provenance (honest labelling):\n *  - `en-female-1-neutral` is the ONE AUTHENTIC OuteAI voice — the OuteTTS v3 bundled\n *    default speaker (default_speakers/en-female-1-neutral), copied verbatim.\n *  - `gerbil-voice-2..6` are DERIVED voices: distinct timbres produced by running\n *    pitch/tempo-varied renderings of that authentic clip back through the REAL\n *    DAC.speech encode + prosody pipeline (scripts/engine/stage-outetts-assets.py).\n *    They are genuine DAC-encoded OuteSpeakers, but they are NOT OuteAI's own\n *    `en-male-*` voices, so they are not labelled as such. (Earlier builds mislabelled\n *    them as `en-male-2-neutral` etc.; renamed for honesty.)\n */\nexport const OUTETTS_PRESET_VOICES = [\n  \"en-female-1-neutral\",\n  \"gerbil-voice-2\",\n  \"gerbil-voice-3\",\n  \"gerbil-voice-4\",\n  \"gerbil-voice-5\",\n  \"gerbil-voice-6\",\n] as const;\n\n/** Resolve the model repo for a set of options, falling back to the defaults. */\nexport function resolveDefaultRepo(opts: {\n  repo?: string;\n  embedding?: boolean;\n  enableVision?: boolean;\n}): string {\n  if (opts.repo) return opts.repo;\n  if (opts.embedding) return DEFAULT_MODELS.embedding;\n  if (opts.enableVision) return DEFAULT_MODELS.vision;\n  return DEFAULT_MODELS.text;\n}\n"],"mappings":";;;;;;AAKA,MAAa,iBAAiB;CAE5B,MAAM;CAEN,QAAQ;CAER,WAAW;CAEX,KAAK;CAKL,SAAS;CAMT,WAAW;CAEX,KAAK;CACN;;;;;;;;AASD,MAAM,qBACJ;AACF,MAAa,iBAAiB;CAE5B,iBAAiB,GAAG,mBAAmB;CAEvC,iBAAiB,GAAG,mBAAmB;CAEvC,cAAc;CACf;;;;;;;;;;;;;;AAeD,MAAa,wBAAwB;CACnC;CACA;CACA;CACA;CACA;CACA;CACD;;AAGD,SAAgB,mBAAmB,MAIxB;AACT,KAAI,KAAK,KAAM,QAAO,KAAK;AAC3B,KAAI,KAAK,UAAW,QAAO,eAAe;AAC1C,KAAI,KAAK,aAAc,QAAO,eAAe;AAC7C,QAAO,eAAe"}
+{"version":3,"file":"defaults-DfGx4d1m.mjs","names":[],"sources":["../src/gpu/defaults.ts"],"sourcesContent":["/**\n * Default model per capability. Kept in its own tiny module (no heavy imports)\n * so the React hooks can resolve defaults without statically pulling in the GPU\n * engine — they import the engine dynamically to stay light.\n */\nexport const DEFAULT_MODELS = {\n  /** Text generation (also the vision-capable checkpoint). */\n  text: \"mlx-community/Qwen3.5-0.8B-4bit\",\n  /** Image understanding — same checkpoint, vision tower built on demand. */\n  vision: \"mlx-community/Qwen3.5-0.8B-4bit\",\n  /** Text embeddings. */\n  embedding: \"mlx-community/embeddinggemma-300m-4bit\",\n  /** Text-to-speech (Apache-2.0; LFM2-350M backbone + NanoCodec). Small/fast default. */\n  tts: \"nineninesix/kani-tts-450m-0.2-ft\",\n  /**\n   * Selectable many-voices TTS (OuteTTS-1.0-0.6B: Qwen3-0.6B codec-LM + DAC.speech).\n   * Pass this as the `repo`/`model` to engine.speak()/useTTS to route to OuteTTS.\n   */\n  ttsOute: \"OuteAI/OuteTTS-1.0-0.6B\",\n  /**\n   * Selectable describe-your-voice TTS (Parler-TTS-mini-v1: Flan-T5 encoder +\n   * cross-attending decoder LM + dac_44khz). Pass this as the `repo`/`model` to\n   * engine.speak()/useTTS with a `describeVoice` natural-language description.\n   */\n  ttsParler: \"parler-tts/parler-tts-mini-v1\",\n  /** Speech-to-text. */\n  stt: \"UsefulSensors/moonshine-base\",\n} as const;\n\n/**\n * OuteTTS auxiliary assets. The DAC.speech codec ships only `.pth` (non-safetensors),\n * so the folded weights (prep-dac-speech.py output) and the preset speaker JSONs are\n * hosted on the Cloudflare R2 model mirror (same public host the engine reads via\n * GERBIL_MODEL_BASE — range + CORS, free egress, no redirect). Override via the OuteTTS\n * options when self-hosting.\n */\nconst OUTETTS_ASSET_BASE =\n  \"https://pub-0522ecec25fc412bb35a523f3e3f63c6.r2.dev/tryhamster/gerbil-outetts-assets\";\nexport const OUTETTS_ASSETS = {\n  /** Folded DAC.speech.v1.0 encoder+decoder weights (prep-dac-speech.py output). */\n  codecWeightsURL: `${OUTETTS_ASSET_BASE}/dac-speech-folded.safetensors`,\n  /** Base URL for the preset speaker JSONs (`{name}.json`). */\n  speakersBaseURL: `${OUTETTS_ASSET_BASE}/speakers`,\n  /** The default preset voice. */\n  defaultVoice: \"en-female-1-neutral\",\n} as const;\n\n/**\n * Bundled OuteTTS preset voices.\n *\n * Provenance (honest labelling):\n *  - `en-female-1-neutral` is the ONE AUTHENTIC OuteAI voice — the OuteTTS v3 bundled\n *    default speaker (default_speakers/en-female-1-neutral), copied verbatim.\n *  - `gerbil-voice-2..6` are DERIVED voices: distinct timbres produced by running\n *    pitch/tempo-varied renderings of that authentic clip back through the REAL\n *    DAC.speech encode + prosody pipeline (scripts/engine/stage-outetts-assets.py).\n *    They are genuine DAC-encoded OuteSpeakers, but they are NOT OuteAI's own\n *    `en-male-*` voices, so they are not labelled as such. (Earlier builds mislabelled\n *    them as `en-male-2-neutral` etc.; renamed for honesty.)\n */\nexport const OUTETTS_PRESET_VOICES = [\n  \"en-female-1-neutral\",\n  \"gerbil-voice-2\",\n  \"gerbil-voice-3\",\n  \"gerbil-voice-4\",\n  \"gerbil-voice-5\",\n  \"gerbil-voice-6\",\n] as const;\n\n/**\n * True when `repo`/`architecture` identify a text-to-speech checkpoint that must\n * run through a dedicated native TTS engine (KaniTTS, OuteTTS, or Parler-TTS)\n * rather than the single-graph base text loader. Mirrors the routing used by\n * `WebGPUEngine.speak()` / `Gerbil.ensureNativeTTSEngine`:\n *   - Parler  → encoder-decoder (`ParlerTTSForConditionalGeneration`)\n *   - OuteTTS → Qwen3-0.6B codec-LM\n *   - Kani    → conv-backbone codec-LM (`KaniTTS2ForCausalLM`, or the Apache\n *               default which is a plain `Lfm2ForCausalLM` — hence the repo-name\n *               gate; the architecture string alone does not identify it).\n * At `create()` time only the repo is known (the architecture is discovered\n * during load), so the repo-name signals below cover all three DEFAULT_MODELS\n * and conventionally-named TTS repos; the architecture arg is honored when a\n * caller already has it.\n */\nexport function isTTSRepo(repo?: string, architecture?: string): boolean {\n  const r = repo ?? \"\";\n  const a = architecture ?? \"\";\n  return (\n    r === DEFAULT_MODELS.tts ||\n    r === DEFAULT_MODELS.ttsOute ||\n    r === DEFAULT_MODELS.ttsParler ||\n    /parler/i.test(r) ||\n    /parler/i.test(a) ||\n    /outetts/i.test(r) ||\n    /oute/i.test(a) ||\n    a === \"KaniTTS2ForCausalLM\" ||\n    /kani-tts/i.test(r) ||\n    /\\btts\\b/i.test(r)\n  );\n}\n\n/** Resolve the model repo for a set of options, falling back to the defaults. */\nexport function resolveDefaultRepo(opts: {\n  repo?: string;\n  embedding?: boolean;\n  enableVision?: boolean;\n}): string {\n  if (opts.repo) return opts.repo;\n  if (opts.embedding) return DEFAULT_MODELS.embedding;\n  if (opts.enableVision) return DEFAULT_MODELS.vision;\n  return DEFAULT_MODELS.text;\n}\n"],"mappings":";;;;;;AAKA,MAAa,iBAAiB;CAE5B,MAAM;CAEN,QAAQ;CAER,WAAW;CAEX,KAAK;CAKL,SAAS;CAMT,WAAW;CAEX,KAAK;CACN;;;;;;;;AASD,MAAM,qBACJ;AACF,MAAa,iBAAiB;CAE5B,iBAAiB,GAAG,mBAAmB;CAEvC,iBAAiB,GAAG,mBAAmB;CAEvC,cAAc;CACf;;;;;;;;;;;;;;AAeD,MAAa,wBAAwB;CACnC;CACA;CACA;CACA;CACA;CACA;CACD;;;;;;;;;;;;;;;;AAiBD,SAAgB,UAAU,MAAe,cAAgC;CACvE,MAAM,IAAI,QAAQ;CAClB,MAAM,IAAI,gBAAgB;AAC1B,QACE,MAAM,eAAe,OACrB,MAAM,eAAe,WACrB,MAAM,eAAe,aACrB,UAAU,KAAK,EAAE,IACjB,UAAU,KAAK,EAAE,IACjB,WAAW,KAAK,EAAE,IAClB,QAAQ,KAAK,EAAE,IACf,MAAM,yBACN,YAAY,KAAK,EAAE,IACnB,WAAW,KAAK,EAAE;;;AAKtB,SAAgB,mBAAmB,MAIxB;AACT,KAAI,KAAK,KAAM,QAAO,KAAK;AAC3B,KAAI,KAAK,UAAW,QAAO,eAAe;AAC1C,KAAI,KAAK,aAAc,QAAO,eAAe;AAC7C,QAAO,eAAe"}

package/dist/frameworks/express.mjs

@@ -1,5 +1,5 @@
 import { t as __require } from "../chunk-BeNUM--j.mjs";
-import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-BpYemKEH.mjs";
+import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-CfNSSpgW.mjs";
 import "../utils-CZBZ8dgR.mjs";
 
 //#region src/frameworks/express.ts

package/dist/frameworks/fastify.mjs

@@ -1,4 +1,4 @@
-import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-BpYemKEH.mjs";
+import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-CfNSSpgW.mjs";
 import "../utils-CZBZ8dgR.mjs";
 
 //#region src/frameworks/fastify.ts

package/dist/frameworks/hono.mjs

@@ -1,4 +1,4 @@
-import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-BpYemKEH.mjs";
+import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-CfNSSpgW.mjs";
 import "../utils-CZBZ8dgR.mjs";
 
 //#region src/frameworks/hono.ts

package/dist/frameworks/next.d.mts

@@ -1,7 +1,7 @@
 import "../outetts-CAL3_K3j.mjs";
 import { d as GerbilConfig } from "../types-BzbBDoaP.mjs";
-import "../index-DyqcKFfQ.mjs";
-import { t as Gerbil } from "../gerbil-6E0XH_8s.mjs";
+import "../index-t_hY-XWW.mjs";
+import { t as Gerbil } from "../gerbil-Cs8L0iVm.mjs";
 
 //#region src/frameworks/next.d.ts
 

package/dist/frameworks/next.mjs

@@ -1,4 +1,4 @@
-import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-BpYemKEH.mjs";
+import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-CfNSSpgW.mjs";
 import "../utils-CZBZ8dgR.mjs";
 
 //#region src/frameworks/next.ts

package/dist/frameworks/trpc.mjs

@@ -1,4 +1,4 @@
-import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-BpYemKEH.mjs";
+import { r as DEFAULT_MODEL, t as Gerbil } from "../gerbil-CfNSSpgW.mjs";
 import "../utils-CZBZ8dgR.mjs";
 import { z } from "zod";
 

package/dist/{gerbil-BpYemKEH.mjs → gerbil-CfNSSpgW.mjs}

@@ -1619,7 +1619,7 @@ The JSON must conform to this schema: ${JSON.stringify(zodToJsonSchema(schema))}
 	/** Lazily build (or reuse) the native MoonshineSTT engine (default repo). */
 	async ensureNativeSTT() {
 		if (!this.nativeSTT) {
-			const { MoonshineSTT } = await import("./moonshine-stt-CcVt4Vdd.mjs");
+			const { MoonshineSTT } = await import("./moonshine-stt-CZ00EdnB.mjs");
 			this.nativeSTT = await MoonshineSTT.create();
 		}
 		return this.nativeSTT;
@@ -1786,4 +1786,4 @@ The JSON must conform to this schema: ${JSON.stringify(zodToJsonSchema(schema))}
 
 //#endregion
 export { resolveModel as a, configureGlobalCache as c, listBuiltinModels as i, getGlobalCache as l, BUILTIN_MODELS as n, ResponseCache as o, DEFAULT_MODEL as r, clearGlobalCache as s, Gerbil as t };
-//# sourceMappingURL=gerbil-BpYemKEH.mjs.map
+//# sourceMappingURL=gerbil-CfNSSpgW.mjs.map