@glissade/narrate 0.23.0-pre.2 → 0.23.0-pre.4

package/dist/providers.d.ts

@@ -20,10 +20,12 @@ interface ZhG2p {
 
 interface EnG2p {
   readonly id: string;
-  /** the g2p identity, folded into kokoroProvider.version() → the cache key */
-  version(): string;
-  /** English text → a misaki[en] phoneme string (US English) */
-  phonemize(text: string): string;
+  /** the g2p identity, folded into kokoroProvider.version() → the cache key.
+   *  `british` ⇒ GB dialect, a DISTINCT cache entry from US (default false). */
+  version(british?: boolean): string;
+  /** English text → a misaki[en] phoneme string. `british` ⇒ GB (`bf_`/`bm_`
+   *  voices); default US (`af_`/`am_`). */
+  phonemize(text: string, british?: boolean): string;
 }
 /**
  * Fork B: shell out to a pinned Python misaki[en]. Mirrors {@link misakiZhG2p} —
@@ -164,6 +166,9 @@ interface ResolvedBlend {
   entries: readonly (readonly [string, number])[];
   /** 'zh' (all zf_/zm_) or 'en' (all other / English) — the g2p front-end. */
   language: 'zh' | 'en';
+  /** For English: 'gb' (all bf_/bm_) or 'us' (all af_/am_ / other) — the misaki[en]
+   *  dialect (`british` flag). Undefined for 'zh'. A mixed US+GB blend throws. */
+  dialect?: 'us' | 'gb';
 }
 /**
  * Validate + normalize a `VoiceBlend`: ≥2 entries, finite positive weights,
@@ -212,6 +217,8 @@ declare function loadKokoroVoiceData(entry: string, name: string): Float32Array;
 declare function kokoroGenerateFromBlend(tts: KokoroModel, Tensor: TensorCtor, inputIds: KokoroInputIds, blended: Float32Array, speed: number): Promise<KokoroAudio>;
 /** A z* (zf_/zm_) kokoro voice routes through misaki[zh] g2p (not espeak cmn). */
 declare function isKokoroChineseVoice(voice: string): boolean;
+/** A b* (bf_/bm_) kokoro voice is BRITISH English → misaki[en] british=True. */
+declare function isKokoroBritishVoice(voice: string): boolean;
 declare function kokoroProvider(opts?: {
   model?: string;
   voice?: VoiceSpec;
@@ -350,4 +357,4 @@ declare function synthesizeScript(scriptPath: string, opts?: SynthesizeOptions):
 /** Resolve `<scene>.narration.json` for a scene-module path (or accept the script itself). */
 declare function scriptPathFor(input: string): string;
 //#endregion
-export { AlignRequest, Aligner, BLEND_SPEC_VERSION, KokoroDtype, ResolvedBlend, SynthesizeOptions, SynthesizeResult, TtsProvider, TtsRequest, TtsResult, VoskAlignWord, alignerById, blendIdentity, blendStyleVectors, cacheKey, espeakProvider, fakeProvider, floatToWav, heuristicAligner, heuristicWords, interpolateMissing, isKokoroChineseVoice, kokoroGenerateFromBlend, kokoroProvider, loadKokoroVoiceData, mapAsrToScript, openaiProvider, piperProvider, providerById, requireStringVoice, resolveBlend, resolvePiperVoice, scriptPathFor, stderrTail, synthesizeScript, voskAligner, wavDuration };
+export { AlignRequest, Aligner, BLEND_SPEC_VERSION, KokoroDtype, ResolvedBlend, SynthesizeOptions, SynthesizeResult, TtsProvider, TtsRequest, TtsResult, VoskAlignWord, alignerById, blendIdentity, blendStyleVectors, cacheKey, espeakProvider, fakeProvider, floatToWav, heuristicAligner, heuristicWords, interpolateMissing, isKokoroBritishVoice, isKokoroChineseVoice, kokoroGenerateFromBlend, kokoroProvider, loadKokoroVoiceData, mapAsrToScript, openaiProvider, piperProvider, providerById, requireStringVoice, resolveBlend, resolvePiperVoice, scriptPathFor, stderrTail, synthesizeScript, voskAligner, wavDuration };

package/dist/providers.js

@@ -176,9 +176,12 @@ function misakiZhG2p(opts = {}) {
 * model + espeak-ng); each missing piece exits with a distinct code → an
 * actionable hint.
 *
-* DIALECT: US English (`british=False`) — the `af_`/`am_` (American) voices.
-* GB-voice (`bf_`/`bm_`) blends are a documented follow-up; the dialect folds
-* into `version()` so US and GB never share a cache entry.
+* DIALECT: US (`british=False`, the `af_`/`am_` voices) OR GB (`british=True`,
+* `bf_`/`bm_`) — `phonemize(text, british)` and `version(british)` take the
+* dialect per call (the caller derives it from the blend's voice prefixes via
+* `resolveBlend`). The dialect folds into `version()` (`dialect=us`/`gb`) so US
+* and GB never share a cache entry. A mixed US+GB blend is rejected upstream
+* (different espeak phoneme front-ends, like mixed languages).
 *
 * DETERMINISM: g2p runs at PREPARE time, never at render. Its identity — a PURE
 * pin string `misaki=<PIN> map=<MAP> dialect=us` (NO Python, no spawn) — folds
@@ -238,13 +241,15 @@ def _check_pins():
         sys.exit(96)
 
 def _g2p():
+    # dialect: --british ⇒ GB English (bf_/bm_ voices); default US (af_/am_)
+    british = "--british" in sys.argv[1:]
     try:
         from misaki import espeak
-        fallback = espeak.EspeakFallback(british=False)
+        fallback = espeak.EspeakFallback(british=british)
     except Exception:
         sys.exit(95)
     try:
-        g = en.G2P(trf=False, british=False, fallback=fallback)
+        g = en.G2P(trf=False, british=british, fallback=fallback)
     except OSError:
         # spaCy en_core_web_sm model not downloaded
         sys.exit(94)
@@ -281,14 +286,16 @@ function misakiEnG2p(opts = {}) {
 	const python = opts.python ?? process.env["MISAKI_PYTHON"] ?? DEFAULT_PYTHON;
 	return {
 		id: "misaki-en",
-		version: () => `misaki-en misaki=${MISAKI_PIN} map=${EN_PHONEME_MAP_VERSION} dialect=us`,
-		phonemize: (text) => {
-			const r = spawnSync(python, [
+		version: (british = false) => `misaki-en misaki=${MISAKI_PIN} map=${EN_PHONEME_MAP_VERSION} dialect=${british ? "gb" : "us"}`,
+		phonemize: (text, british = false) => {
+			const args = [
 				"-c",
 				EN_G2P_PY,
 				"--pins",
 				MISAKI_PIN
-			], {
+			];
+			if (british) args.push("--british");
+			const r = spawnSync(python, args, {
 				input: text,
 				maxBuffer: 8 * 1024 * 1024
 			});
@@ -611,9 +618,22 @@ function resolveBlend(spec) {
 		const en = raw.filter((_, i) => !chinese[i]).map(([n]) => n);
 		throw new NarrationError(`voice blend mixes languages — Chinese [${zh.join(", ")}] with non-Chinese [${en.join(", ")}]; a blend must share one g2p front-end (all zf_/zm_ OR all English)`);
 	}
+	let dialect;
+	if (allEn) {
+		const british = raw.map(([name]) => isKokoroBritishVoice(name));
+		const allGb = british.every(Boolean);
+		const allUs = british.every((b) => !b);
+		if (!allGb && !allUs) {
+			const gb = raw.filter((_, i) => british[i]).map(([n]) => n);
+			const us = raw.filter((_, i) => !british[i]).map(([n]) => n);
+			throw new NarrationError(`voice blend mixes English dialects — British [${gb.join(", ")}] with American [${us.join(", ")}]; a blend must share one espeak front-end (all bf_/bm_ OR all af_/am_)`);
+		}
+		dialect = allGb ? "gb" : "us";
+	}
 	return {
 		entries: raw.map(([name, weight]) => [name, weight / sum]),
-		language: allZh ? "zh" : "en"
+		language: allZh ? "zh" : "en",
+		...dialect !== void 0 ? { dialect } : {}
 	};
 }
 /**
@@ -624,8 +644,8 @@ function resolveBlend(spec) {
 * segment cache, EXACTLY the 0.15 g2p-identity pattern.
 */
 function blendIdentity(spec) {
-	const { entries, language } = resolveBlend(spec);
-	return `blend=[${entries.map(([name, w]) => `${name}:${w.toFixed(6)}`).join(",")} lang=${language} v${BLEND_SPEC_VERSION}]`;
+	const { entries, language, dialect } = resolveBlend(spec);
+	return `blend=[${entries.map(([name, w]) => `${name}:${w.toFixed(6)}`).join(",")} lang=${language}${dialect !== void 0 ? ` dialect=${dialect}` : ""} v${BLEND_SPEC_VERSION}]`;
 }
 /**
 * Compute the normalized weighted sum of base-voice style tensors into one
@@ -756,6 +776,10 @@ async function kokoroGenerateFromBlend(tts, Tensor, inputIds, blended, speed) {
 function isKokoroChineseVoice(voice) {
 	return voice.startsWith("zf_") || voice.startsWith("zm_");
 }
+/** A b* (bf_/bm_) kokoro voice is BRITISH English → misaki[en] british=True. */
+function isKokoroBritishVoice(voice) {
+	return voice.startsWith("bf_") || voice.startsWith("bm_");
+}
 function kokoroProvider(opts = {}) {
 	const modelId = opts.model ?? KOKORO_MODEL;
 	const dtype = opts.dtype ?? "q8";
@@ -792,7 +816,8 @@ function kokoroProvider(opts = {}) {
 			let blendSuffix = "";
 			if (isVoiceBlend(opts.voice)) {
 				blendSuffix = ` ${blendIdentity(opts.voice)}`;
-				if (resolveBlend(opts.voice).language === "en") blendSuffix += ` en-g2p=[${enG2p.version()}]`;
+				const resolvedV = resolveBlend(opts.voice);
+				if (resolvedV.language === "en") blendSuffix += ` en-g2p=[${enG2p.version(resolvedV.dialect === "gb")}]`;
 			}
 			const v = `kokoro-js ${version} ${basename(modelId)} dtype=${dtype} g2p=[${zhG2p.version()}]${blendSuffix}`;
 			return Promise.resolve(v);
@@ -806,14 +831,14 @@ function kokoroProvider(opts = {}) {
 				const blended = blendStyleVectors(resolved.entries, (name) => loadKokoroVoiceData(entry, name));
 				const Tensor = await loadTensorCtor(entry);
 				const recipe = resolved.entries.map(([n, w]) => `${n}*${w.toFixed(4)}`).join(" + ");
-				console.log(`[narrate] kokoro blended voice (Apache-2.0, derived) lang=${resolved.language}: ${recipe}`);
+				console.log(`[narrate] kokoro blended voice (Apache-2.0, derived) lang=${resolved.language}${resolved.dialect !== void 0 ? `/${resolved.dialect}` : ""}: ${recipe}`);
 				let inputIds;
 				try {
 					if (resolved.language === "zh") {
 						const phonemes = zhG2p.phonemize(req.text);
 						inputIds = tts.tokenizer(phonemes, { truncation: true }).input_ids;
 					} else {
-						const phonemes = enG2p.phonemize(req.text);
+						const phonemes = enG2p.phonemize(req.text, resolved.dialect === "gb");
 						inputIds = tts.tokenizer(phonemes, { truncation: true }).input_ids;
 					}
 					const audio = await kokoroGenerateFromBlend(tts, Tensor, inputIds, blended, speed ?? 1);
@@ -1209,4 +1234,4 @@ function scriptPathFor(input) {
 	return candidate;
 }
 //#endregion
-export { BLEND_SPEC_VERSION, alignerById, blendIdentity, blendStyleVectors, cacheKey, espeakProvider, fakeProvider, floatToWav, heuristicAligner, heuristicWords, interpolateMissing, isKokoroChineseVoice, kokoroGenerateFromBlend, kokoroProvider, loadKokoroVoiceData, mapAsrToScript, openaiProvider, piperProvider, providerById, requireStringVoice, resolveBlend, resolvePiperVoice, scriptPathFor, stderrTail, synthesizeScript, voskAligner, wavDuration };
+export { BLEND_SPEC_VERSION, alignerById, blendIdentity, blendStyleVectors, cacheKey, espeakProvider, fakeProvider, floatToWav, heuristicAligner, heuristicWords, interpolateMissing, isKokoroBritishVoice, isKokoroChineseVoice, kokoroGenerateFromBlend, kokoroProvider, loadKokoroVoiceData, mapAsrToScript, openaiProvider, piperProvider, providerById, requireStringVoice, resolveBlend, resolvePiperVoice, scriptPathFor, stderrTail, synthesizeScript, voskAligner, wavDuration };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@glissade/narrate",
-  "version": "0.23.0-pre.2",
+  "version": "0.23.0-pre.4",
   "description": "glissade narration + captions: TTS at prepare time (gs narrate), deterministic caching, narration-anchored timeline beats, and captions as plain tracks. Render stays offline.",
   "license": "Apache-2.0",
   "engines": {
@@ -22,8 +22,8 @@
     "dist"
   ],
   "dependencies": {
-    "@glissade/core": "0.23.0-pre.2",
-    "@glissade/scene": "0.23.0-pre.2"
+    "@glissade/core": "0.23.0-pre.4",
+    "@glissade/scene": "0.23.0-pre.4"
   },
   "peerDependencies": {
     "kokoro-js": "^1.2.0"