CTC forced alignment for word timings (supersedes #61, #65)

lib/alignModels.ts

@@ -0,0 +1,56 @@
+import type { TranscriptLanguage } from "./languages";
+import type { CtcNormalizeMode } from "./forcedAlign";
+
+/** CTC acoustic model used to force-align a transcript to audio. */
+export interface AlignModelInfo {
+  /** Hugging Face / transformers.js model id. */
+  id: string;
+  /** How transcript text is folded into this model's character vocabulary. */
+  normalize: CtcNormalizeMode;
+}
+
+/**
+ * Language → CTC aligner map (WhisperX-style).
+ *
+ * English keeps the small Librispeech wav2vec2. European languages share the
+ * multilingual MMS forced-aligner (Latin a–z after accent folding). Chinese
+ * uses an XLS-R CTC model whose vocab includes Han characters. Languages with
+ * no entry skip CTC and keep the VAD/envelope heuristic only.
+ */
+export const ALIGN_MODELS: Partial<
+  Record<TranscriptLanguage, AlignModelInfo>
+> = {
+  en: {
+    id: "Xenova/wav2vec2-base-960h",
+    normalize: "latin-upper",
+  },
+  es: {
+    id: "onnx-community/mms-300m-1130-forced-aligner-ONNX",
+    normalize: "latin-lower",
+  },
+  fr: {
+    id: "onnx-community/mms-300m-1130-forced-aligner-ONNX",
+    normalize: "latin-lower",
+  },
+  de: {
+    id: "onnx-community/mms-300m-1130-forced-aligner-ONNX",
+    normalize: "latin-lower",
+  },
+  zh: {
+    id: "onnx-community/wav2vec2-large-xlsr-53-chinese-zh-cn-ONNX",
+    normalize: "cjk",
+  },
+};
+
+/** Align model for a transcript language, or null when CTC is unavailable. */
+export function alignModelFor(
+  language: TranscriptLanguage | string | undefined
+): AlignModelInfo | null {
+  if (!language) return null;
+  return ALIGN_MODELS[language as TranscriptLanguage] ?? null;
+}
+
+/** Unique Hugging Face ids referenced by {@link ALIGN_MODELS}. */
+export function uniqueAlignModelIds(): string[] {
+  return [...new Set(Object.values(ALIGN_MODELS).map((m) => m.id))];
+}