@glissade/narrate 0.5.0-pre.0 → 0.5.0-pre.1

package/dist/providers.d.ts

@@ -111,23 +111,29 @@ declare function mapAsrToScript(timed: {
   start: number;
   end: number;
 }[];
-interface WavMono {
-  /** mono samples in [-1, 1] */
-  samples: Float32Array;
-  sampleRate: number;
+/** one word from vosk-align's JSON output */
+interface VoskAlignWord {
+  word: string;
+  start: number;
+  end: number;
+  conf?: number;
 }
-/** Decode a 16-bit PCM RIFF/WAV to mono float samples (channels averaged). */
-declare function decodeWavMono(wav: Buffer): WavMono;
-/** Linear-resample mono float to a 16 kHz int16 LE PCM buffer (Vosk's input). */
-declare function resampleTo16kPcm(input: WavMono): Buffer;
 /**
- * Word timings via Vosk (alphacephei) — offline, Apache-2.0, ~50 MB model, a
- * real Node binding (no Python, no Docker, no multi-GB download). `vosk` is an
- * OPTIONAL peer: install it (`npm i vosk`) and point at a model
- * (`opts.model` / `VOSK_MODEL`) only if you use this aligner. ASR words are
- * mapped onto the script tokens by `mapAsrToScript`.
+ * Word timings via Vosk (alphacephei) — offline ASR, Apache-2.0. Shells out to
+ * a `vosk-align` command (the Python `vosk` binding + ffmpeg — deliberately NOT
+ * the npm `vosk` package, whose `ffi-napi` native build is broken on modern
+ * Node). The command reads any audio and writes
+ *   { "words": [ { "word", "start", "end", "conf"? }, … ] }
+ * to stdout; its recognized words are LCS-mapped onto the script tokens by
+ * `mapAsrToScript`, so mis-recognitions (e.g. an unknown proper noun) just
+ * interpolate cleanly between the words around them.
+ *
+ * Provide the command via `opts.command` / `VOSK_ALIGN` (default `vosk-align`);
+ * the model is the command's own concern (its default, or `--model`/VOSK_MODEL),
+ * passed through with `opts.model`.
  */
 declare function voskAligner(opts?: {
+  command?: string;
   model?: string;
 }): Aligner;
 /** Resolve an aligner id; 'none' disables alignment (word-less segments). */
@@ -170,4 +176,4 @@ 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, SynthesizeOptions, SynthesizeResult, TtsProvider, TtsRequest, TtsResult, alignerById, cacheKey, decodeWavMono, espeakProvider, fakeProvider, heuristicAligner, heuristicWords, interpolateMissing, mapAsrToScript, openaiProvider, piperProvider, providerById, resampleTo16kPcm, scriptPathFor, synthesizeScript, voskAligner, wavDuration };
+export { AlignRequest, Aligner, SynthesizeOptions, SynthesizeResult, TtsProvider, TtsRequest, TtsResult, VoskAlignWord, alignerById, cacheKey, espeakProvider, fakeProvider, heuristicAligner, heuristicWords, interpolateMissing, mapAsrToScript, openaiProvider, piperProvider, providerById, scriptPathFor, synthesizeScript, voskAligner, wavDuration };

package/dist/providers.js

@@ -144,8 +144,12 @@ function piperProvider(opts = {}) {
 		id: "piper",
 		version: () => {
 			const r = spawnSync("piper", ["--version"], { encoding: "utf8" });
-			if (r.status !== 0) throw new NarrationError("piper not found on PATH — install rhasspy/piper, or use --provider fake/espeak/openai");
-			const v = (r.stdout.trim() || r.stderr.trim() || "piper").split("\n")[0];
+			if (r.error) {
+				if (r.error.code === "ENOENT") throw new NarrationError("piper not found on PATH — `pip install piper-tts` (or the standalone rhasspy/piper), or use --provider fake/espeak/openai");
+				throw new NarrationError(`could not run piper: ${r.error.message}`);
+			}
+			const m = /\b\d+\.\d+\.\d+\b/.exec(r.stdout ?? "");
+			const v = m ? `piper ${m[0]}` : "piper (version unknown)";
 			return Promise.resolve(opts.model ? `${v} ${basename(opts.model)}` : v);
 		},
 		synthesize: (req) => {
@@ -299,100 +303,51 @@ function mapAsrToScript(timed, scriptText) {
 		};
 	}));
 }
-/** Decode a 16-bit PCM RIFF/WAV to mono float samples (channels averaged). */
-function decodeWavMono(wav) {
-	if (wav.length < 44 || wav.toString("ascii", 0, 4) !== "RIFF" || wav.toString("ascii", 8, 12) !== "WAVE") throw new NarrationError("not a RIFF/WAVE file");
-	let channels = 1;
-	let sampleRate = 16e3;
-	let bits = 16;
-	let dataOffset = -1;
-	let dataSize = 0;
-	let offset = 12;
-	while (offset + 8 <= wav.length) {
-		const id = wav.toString("ascii", offset, offset + 4);
-		const size = wav.readUInt32LE(offset + 4);
-		if (id === "fmt ") {
-			channels = wav.readUInt16LE(offset + 10);
-			sampleRate = wav.readUInt32LE(offset + 12);
-			bits = wav.readUInt16LE(offset + 22);
-		} else if (id === "data") {
-			dataOffset = offset + 8;
-			dataSize = size;
-		}
-		offset += 8 + size + size % 2;
-	}
-	if (bits !== 16) throw new NarrationError(`only 16-bit PCM WAV is supported (got ${bits}-bit)`);
-	if (dataOffset < 0) throw new NarrationError("WAV has no data chunk");
-	const frames = Math.floor(dataSize / 2 / Math.max(1, channels));
-	const samples = new Float32Array(frames);
-	for (let f = 0; f < frames; f++) {
-		let acc = 0;
-		for (let c = 0; c < channels; c++) acc += wav.readInt16LE(dataOffset + (f * channels + c) * 2);
-		samples[f] = acc / channels / 32768;
-	}
-	return {
-		samples,
-		sampleRate
-	};
-}
-/** Linear-resample mono float to a 16 kHz int16 LE PCM buffer (Vosk's input). */
-function resampleTo16kPcm(input) {
-	const ratio = input.sampleRate / 16e3;
-	const outLen = Math.max(1, Math.round(input.samples.length / ratio));
-	const out = Buffer.alloc(outLen * 2);
-	for (let i = 0; i < outLen; i++) {
-		const src = i * ratio;
-		const j = Math.floor(src);
-		const frac = src - j;
-		const a = input.samples[j] ?? 0;
-		const b = input.samples[j + 1] ?? a;
-		const v = Math.max(-1, Math.min(1, a + (b - a) * frac));
-		out.writeInt16LE(Math.round(v * 32767), i * 2);
-	}
-	return out;
-}
 /**
-* Word timings via Vosk (alphacephei) — offline, Apache-2.0, ~50 MB model, a
-* real Node binding (no Python, no Docker, no multi-GB download). `vosk` is an
-* OPTIONAL peer: install it (`npm i vosk`) and point at a model
-* (`opts.model` / `VOSK_MODEL`) only if you use this aligner. ASR words are
-* mapped onto the script tokens by `mapAsrToScript`.
+* Word timings via Vosk (alphacephei) — offline ASR, Apache-2.0. Shells out to
+* a `vosk-align` command (the Python `vosk` binding + ffmpeg — deliberately NOT
+* the npm `vosk` package, whose `ffi-napi` native build is broken on modern
+* Node). The command reads any audio and writes
+*   { "words": [ { "word", "start", "end", "conf"? }, … ] }
+* to stdout; its recognized words are LCS-mapped onto the script tokens by
+* `mapAsrToScript`, so mis-recognitions (e.g. an unknown proper noun) just
+* interpolate cleanly between the words around them.
+*
+* Provide the command via `opts.command` / `VOSK_ALIGN` (default `vosk-align`);
+* the model is the command's own concern (its default, or `--model`/VOSK_MODEL),
+* passed through with `opts.model`.
 */
 function voskAligner(opts = {}) {
-	const modelPath = opts.model ?? process.env["VOSK_MODEL"];
-	let vosk = null;
-	const load = async () => {
-		if (vosk) return vosk;
-		try {
-			vosk = await import("vosk");
-		} catch {
-			throw new NarrationError("vosk is not installed — `npm i vosk` and download a model, or use --align heuristic");
-		}
-		vosk.setLogLevel(-1);
-		return vosk;
-	};
+	const command = opts.command ?? process.env["VOSK_ALIGN"] ?? "vosk-align";
 	return {
 		id: "vosk",
-		version: async () => {
-			if (!modelPath) throw new NarrationError("vosk needs a model — set VOSK_MODEL or pass { model } (alphacephei.com/vosk/models)");
-			if (!existsSync(modelPath)) throw new NarrationError(`vosk model not found at ${modelPath}`);
-			await load();
-			return `vosk:${basename(modelPath)}`;
+		version: () => {
+			const r = spawnSync(command, ["--help"], { encoding: "utf8" });
+			if (r.error) {
+				if (r.error.code === "ENOENT") throw new NarrationError(`'${command}' not found — provide a vosk-align command (Apache-2.0 Vosk + ffmpeg, JSON {words:[{word,start,end}]} on stdout), or use --align heuristic`);
+				throw new NarrationError(`could not run ${command}: ${r.error.message}`);
+			}
+			return Promise.resolve(opts.model ? `vosk ${basename(opts.model)}` : "vosk");
 		},
-		align: async (req) => {
-			const v = await load();
-			const model = new v.Model(modelPath);
-			const rec = new v.Recognizer({
-				model,
-				sampleRate: 16e3
-			});
+		align: (req) => {
+			const tag = createHash("sha256").update(req.text).digest("hex").slice(0, 8);
+			const wavPath = join(tmpdir(), `glissade-vosk-${process.pid}-${tag}.wav`);
 			try {
-				rec.setWords(true);
-				rec.acceptWaveform(resampleTo16kPcm(decodeWavMono(req.wav)));
-				return mapAsrToScript(rec.finalResult().result ?? [], req.text);
+				writeFileSync(wavPath, req.wav);
+				const r = spawnSync(command, [wavPath, ...opts.model ? ["--model", opts.model] : []], {
+					encoding: "utf8",
+					maxBuffer: 64 * 1024 * 1024
+				});
+				if (r.error) throw new NarrationError(`${command} failed to run: ${r.error.message}`);
+				if (r.status !== 0) throw new NarrationError(`${command} failed: ${(r.stderr || "").slice(0, 300)}`);
+				const timed = (JSON.parse(r.stdout).words ?? []).filter((w) => typeof w.start === "number" && typeof w.end === "number").map((w) => ({
+					word: w.word,
+					start: w.start,
+					end: w.end
+				}));
+				return Promise.resolve(mapAsrToScript(timed, req.text));
 			} finally {
-				rec.free();
-				model.free();
+				if (existsSync(wavPath)) unlinkSync(wavPath);
 			}
 		}
 	};
@@ -539,4 +494,4 @@ function scriptPathFor(input) {
 	return candidate;
 }
 //#endregion
-export { alignerById, cacheKey, decodeWavMono, espeakProvider, fakeProvider, heuristicAligner, heuristicWords, interpolateMissing, mapAsrToScript, openaiProvider, piperProvider, providerById, resampleTo16kPcm, scriptPathFor, synthesizeScript, voskAligner, wavDuration };
+export { alignerById, cacheKey, espeakProvider, fakeProvider, heuristicAligner, heuristicWords, interpolateMissing, mapAsrToScript, openaiProvider, piperProvider, providerById, scriptPathFor, synthesizeScript, voskAligner, wavDuration };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@glissade/narrate",
-  "version": "0.5.0-pre.0",
+  "version": "0.5.0-pre.1",
   "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",
   "type": "module",
@@ -19,8 +19,8 @@
     "dist"
   ],
   "dependencies": {
-    "@glissade/core": "0.5.0-pre.0",
-    "@glissade/scene": "0.5.0-pre.0"
+    "@glissade/core": "0.5.0-pre.1",
+    "@glissade/scene": "0.5.0-pre.1"
   },
   "repository": {
     "type": "git",