@broberg/ai-sdk 0.42.1 → 0.44.0

package/dist/index.d.ts

@@ -61,6 +61,24 @@ interface Pronunciation {
     word: string;
     /** Say this instead. Azure `<sub alias>`; ElevenLabs plain substitution. */
     alias?: string;
+    /** Also match INSIDE a hyphenated compound — `AI` in `AI-agenter` (F051.3).
+     *
+     *  Danish puts abbreviations there constantly. cms measured their own articles:
+     *  "AI-agenter" 13 times, "AI-native" 12, 60+ occurrences in all, every one read aloud
+     *  as one mangled word. Christian heard it as "HTLM" and asked why AI was not spelled
+     *  out too — it WAS, just never in a compound.
+     *
+     *  **Per entry, and defaulting to `false`, because the right answer depends on the
+     *  word.** `AI` yes; `mail` no — a rule for `mail` must not fire inside `e-mail`.
+     *  Measured rather than reasoned: `@broberg/speech-dictionary` ships the general
+     *  loosening and pays exactly that price today. A single global answer is wrong for
+     *  one of the two classes whichever way it is set, so the person who wrote the entry
+     *  decides.
+     *
+     *  Default `false` is also the safe direction to be wrong in: flipping it would change
+     *  the audio for every existing consumer IN SOUND, where there is no compile error, no
+     *  red test and no log line to notice it by. */
+    matchInCompounds?: boolean;
     /** IPA phonemes. Azure `<phoneme alphabet="ipa" ph>`. ElevenLabs cannot do this
      *  (it THROWS rather than silently skipping). See the type doc above before reaching
      *  for this on a foreign word. */
@@ -461,6 +479,11 @@ interface DialogueTurn {
 }
 interface DialogueRequest {
     inputs: DialogueTurn[];
+    /** F051.4 — the pronunciation dictionary, applied to EACH line. cms measured the
+     *  gap: a two-host podcast could not get a dictionary at all, so "broberg.ai" was
+     *  said wrong in every episode — they could fix the sponsor read and the host's
+     *  hand-off (both `ai.tts`) and not the conversation, which is 95% of the audio. */
+    pronunciations?: Pronunciation[];
     /** Output container, e.g. "mp3" (default). */
     format?: string;
     spec: TierSpec;
@@ -1819,6 +1842,29 @@ declare const podcastInputSchema: z.ZodObject<{
     }>, "many">;
     voices: z.ZodRecord<z.ZodString, z.ZodString>;
     format: z.ZodOptional<z.ZodString>;
+    /** F051.4 — same field, same semantics as tts. Applied PER LINE: ElevenLabs'
+     *  /text-to-dialogue takes inputs[].text separately, so there is no composed string
+     *  a replacement could run across a speaker boundary in. */
+    pronunciations: z.ZodOptional<z.ZodArray<z.ZodObject<{
+        word: z.ZodString;
+        alias: z.ZodOptional<z.ZodString>;
+        ipa: z.ZodOptional<z.ZodString>;
+        lang: z.ZodOptional<z.ZodString>;
+        /** F051.3 — also match inside a hyphenated compound ("AI" in "AI-agenter"). */
+        matchInCompounds: z.ZodOptional<z.ZodBoolean>;
+    }, "strip", z.ZodTypeAny, {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }, {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }>, "many">>;
 }, "strip", z.ZodTypeAny, {
     script: {
         text: string;
@@ -1839,6 +1885,13 @@ declare const podcastInputSchema: z.ZodObject<{
     })[] | undefined;
     labels?: Record<string, string> | undefined;
     format?: string | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
 }, {
     script: {
         text: string;
@@ -1859,6 +1912,13 @@ declare const podcastInputSchema: z.ZodObject<{
     })[] | undefined;
     labels?: Record<string, string> | undefined;
     format?: string | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
 }>;
 declare const ttsInputSchema: z.ZodObject<{
     tier: z.ZodOptional<z.ZodEnum<["fast", "smart", "powerful", "cheap", "vision", "video", "embedding"]>>;
@@ -1899,13 +1959,17 @@ declare const ttsInputSchema: z.ZodObject<{
         alias: z.ZodOptional<z.ZodString>;
         ipa: z.ZodOptional<z.ZodString>;
         lang: z.ZodOptional<z.ZodString>;
+        /** F051.3 — also match inside a hyphenated compound ("AI" in "AI-agenter"). */
+        matchInCompounds: z.ZodOptional<z.ZodBoolean>;
     }, "strip", z.ZodTypeAny, {
         word: string;
+        matchInCompounds?: boolean | undefined;
         alias?: string | undefined;
         ipa?: string | undefined;
         lang?: string | undefined;
     }, {
         word: string;
+        matchInCompounds?: boolean | undefined;
         alias?: string | undefined;
         ipa?: string | undefined;
         lang?: string | undefined;
@@ -1940,10 +2004,11 @@ declare const ttsInputSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     })[] | undefined;
     labels?: Record<string, string> | undefined;
-    format?: string | undefined;
     lang?: string | undefined;
+    format?: string | undefined;
     pronunciations?: {
         word: string;
+        matchInCompounds?: boolean | undefined;
         alias?: string | undefined;
         ipa?: string | undefined;
         lang?: string | undefined;
@@ -1966,10 +2031,11 @@ declare const ttsInputSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     })[] | undefined;
     labels?: Record<string, string> | undefined;
-    format?: string | undefined;
     lang?: string | undefined;
+    format?: string | undefined;
     pronunciations?: {
         word: string;
+        matchInCompounds?: boolean | undefined;
         alias?: string | undefined;
         ipa?: string | undefined;
         lang?: string | undefined;
@@ -2366,8 +2432,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.42.1";
-declare const SDK_TAG: "@broberg/ai-sdk@0.42.1";
+declare const VERSION: "0.44.0";
+declare const SDK_TAG: "@broberg/ai-sdk@0.44.0";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.

package/dist/index.js

@@ -1557,10 +1557,14 @@ function applyPronunciations(haystack, list, render, escape = (s) => s) {
     if (!byLower.has(k)) byLower.set(k, p);
   }
   const alternation = sorted.map((p) => escapeRegex(escape(p.word))).join("|");
-  const re = new RegExp(`(?<![\\w-])(${alternation})(?![\\w-])`, "gi");
-  return haystack.replace(re, (matched) => {
+  const re = new RegExp(`(?<!\\w)(${alternation})(?!\\w)`, "gi");
+  return haystack.replace(re, (matched, _g1, offset) => {
     const entry = byLower.get(matched.toLowerCase());
-    return entry ? render(entry, matched) : matched;
+    if (!entry) return matched;
+    if (!entry.matchInCompounds) {
+      if (haystack[offset - 1] === "-" || haystack[offset + matched.length] === "-") return matched;
+    }
+    return render(entry, matched);
   });
 }
 
@@ -1613,7 +1617,13 @@ function elevenlabsAdapter(config = {}) {
       headers: { "xi-api-key": key(), "content-type": "application/json", accept: "audio/mpeg" },
       body: JSON.stringify({
         model_id: req.spec.model,
-        inputs: req.inputs.map((t) => ({ text: t.text, voice_id: t.voiceId })),
+        // F051.4 — PER LINE. The endpoint takes each turn's text separately, so the
+        // dictionary applies exactly as it does in tts; composing the script into one
+        // string first would let a replacement run across a speaker boundary.
+        inputs: req.inputs.map((t) => ({
+          text: ttsText({ text: t.text, pronunciations: req.pronunciations }),
+          voice_id: t.voiceId
+        })),
         ...req.format ? { output_format: req.format } : {}
       })
     });
@@ -3051,18 +3061,24 @@ var moderationInputSchema = z.object({
   input: z.union([z.string(), z.array(z.string())]),
   ...callOptions
 });
+var pronunciationSchema = z.object({
+  word: z.string(),
+  alias: z.string().optional(),
+  ipa: z.string().optional(),
+  lang: z.string().optional(),
+  /** F051.3 — also match inside a hyphenated compound ("AI" in "AI-agenter"). */
+  matchInCompounds: z.boolean().optional()
+});
 var podcastInputSchema = z.object({
   script: z.array(z.object({ speaker: z.string(), text: z.string() })).min(1),
   voices: z.record(z.string(), z.string()),
   format: z.string().optional(),
+  /** F051.4 — same field, same semantics as tts. Applied PER LINE: ElevenLabs'
+   *  /text-to-dialogue takes inputs[].text separately, so there is no composed string
+   *  a replacement could run across a speaker boundary in. */
+  pronunciations: z.array(pronunciationSchema).optional(),
   ...callOptions
 });
-var pronunciationSchema = z.object({
-  word: z.string(),
-  alias: z.string().optional(),
-  ipa: z.string().optional(),
-  lang: z.string().optional()
-});
 var ttsInputSchema = z.object({
   text: z.string(),
   voice: z.string(),
@@ -3110,8 +3126,8 @@ var aiConfigSchema = z.object({
 });
 
 // src/version.ts
-var VERSION = "0.42.1";
-var SDK_TAG = "@broberg/ai-sdk@0.42.1";
+var VERSION = "0.44.0";
+var SDK_TAG = "@broberg/ai-sdk@0.44.0";
 
 // src/cost/sinks/upmetrics.ts
 function upmetricsSink(config) {
@@ -3640,7 +3656,7 @@ function createAI(config = {}) {
         invoke: async (spec) => {
           const adapter = pickProvider(spec.provider);
           if (!adapter.dialogue) throw new Error(`createAI: provider "${spec.provider}" does not support podcast/dialogue`);
-          return adapter.dialogue({ inputs, format: input.format, spec });
+          return adapter.dialogue({ inputs, format: input.format, pronunciations: input.pronunciations, spec });
         }
       });
     },