@broberg/ai-sdk 0.43.0 → 0.45.0

package/dist/index.d.ts

@@ -87,6 +87,61 @@ interface Pronunciation {
     lang?: string;
 }
 
+/** One entry of Azure's `[nnnn].word.json`, verbatim field names. */
+interface AzureWordBoundary {
+    Text: string;
+    /** Milliseconds into the audio. */
+    AudioOffset: number;
+    /** Milliseconds. */
+    Duration: number;
+}
+interface WordTiming {
+    /** The spoken word, as Azure reported it. */
+    text: string;
+    startMs: number;
+    endMs: number;
+    /** Character span in the SOURCE text. Several spoken words share one span when they
+     *  came from one substituted manuscript word. */
+    sourceStart: number;
+    sourceEnd: number;
+}
+interface AlignedWordTimings {
+    /** Every word we could place in the source, in audio order. */
+    words: WordTiming[];
+    /** Spoken words we could NOT place, in order, verbatim.
+     *
+     *  Named rather than dropped, and never given a guessed offset. "I could not place
+     *  this" and "it was here" must not be the same answer — the same discipline as
+     *  `rerank.unscored` (F052). A highlighter that treats `words` as complete must look
+     *  here. */
+    unaligned: string[];
+}
+/** A pronunciation entry, narrowed to what alignment needs. */
+interface AliasEntry {
+    word: string;
+    alias?: string;
+}
+/**
+ * Map Azure's spoken word list onto positions in the SOURCE text.
+ *
+ * Two rules carry the whole function:
+ *
+ * **The walk only moves FORWARD.** That is what puts a repeated word on its own
+ * occurrence — "AI er AI" must not place the second on the first. Searching from 0 each
+ * time passes every test with distinct words and fails silently on real prose.
+ *
+ * **A substitution is matched as a RUN, not word by word.** Azure speaks SSML, so
+ * `<sub alias='broberg punktum a i'>broberg.ai</sub>` puts four words in the list where
+ * the manuscript has one. Matching them individually is not merely imprecise, it is
+ * wrong in a way that bites every Danish text: MEASURED on cms's own dictionary entry,
+ * a per-word map made the alias's "i" swallow the next real Danish word "i" in
+ * "broberg.ai i dag". An alias's parts are often ordinary words ("a", "i", "punktum"),
+ * so a lookup table cannot be the mechanism — the run has to be consumed positionally.
+ */
+declare function alignWordTimings(text: string, words: AzureWordBoundary[], opts?: {
+    pronunciations?: AliasEntry[];
+}): AlignedWordTimings;
+
 /** How a call reaches the model. `http` = provider REST API; `subprocess` = local
  *  `claude -p` CLI (Max plan, costUsd 0). */
 type Transport = "http" | "subprocess";
@@ -479,6 +534,11 @@ 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;
@@ -487,6 +547,14 @@ interface PodcastResult {
     /** Episode audio bytes. */
     audio: Uint8Array;
     mimeType: string;
+    /** F055 — per-word timings, present only when `wordTimings` was asked for AND the
+     *  route produces them. Its ABSENCE says "not available here"; an empty `words` array
+     *  would say "the text had no words", which is a different claim.
+     *
+     *  `sourceStart`/`sourceEnd` index the ORIGINAL text you passed, not the SSML we sent
+     *  — Azure reports only audio time, so the link back to the manuscript is derived
+     *  here. `unaligned` names any spoken word that could not be placed. */
+    wordTimings?: AlignedWordTimings;
     usage: Usage;
 }
 
@@ -507,6 +575,16 @@ interface TtsRequest {
      *  into SSML** — the substitution happens adapter-side AFTER the text is escaped, so
      *  `text` can never inject markup, and `alias`/`ipa` are escaped too. */
     pronunciations?: Pronunciation[];
+    /** F055 — ask for per-word timings so a reader can highlight each word as it is
+     *  spoken. Azure only. Absent from the result means "this route does not produce
+     *  them", which is a different answer from an empty list.
+     *
+     *  **This changes the ROUTE, not just the payload.** The real-time endpoint returns
+     *  audio only; word boundaries exist solely on Azure's asynchronous batch synthesis
+     *  API, so the call becomes submit → poll → fetch. Microsoft quotes 10–20s for half
+     *  of all jobs and up to 120s for 95%. Worth it for a pre-generated, cached reading;
+     *  wrong for anything a user is waiting on. */
+    wordTimings?: boolean;
     spec: TierSpec;
 }
 interface BatchRequestItem {
@@ -1317,6 +1395,7 @@ declare const imageInputSchema: z.ZodObject<{
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     seed?: number | undefined;
     purpose?: string | undefined;
+    outputFormat?: "jpeg" | "png" | "webp" | undefined;
     loras?: {
         path: string;
         scale?: number | undefined;
@@ -1338,7 +1417,6 @@ declare const imageInputSchema: z.ZodObject<{
     finetune?: string | undefined;
     finetuneStrength?: number | undefined;
     referenceImages?: (string | Uint8Array<ArrayBuffer>)[] | undefined;
-    outputFormat?: "jpeg" | "png" | "webp" | undefined;
     safetyTolerance?: number | undefined;
     retryOnBlack?: boolean | undefined;
 }, {
@@ -1346,6 +1424,7 @@ declare const imageInputSchema: z.ZodObject<{
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     seed?: number | undefined;
     purpose?: string | undefined;
+    outputFormat?: "jpeg" | "png" | "webp" | undefined;
     loras?: {
         path: string;
         scale?: number | undefined;
@@ -1367,7 +1446,6 @@ declare const imageInputSchema: z.ZodObject<{
     finetune?: string | undefined;
     finetuneStrength?: number | undefined;
     referenceImages?: (string | Uint8Array<ArrayBuffer>)[] | undefined;
-    outputFormat?: "jpeg" | "png" | "webp" | undefined;
     safetyTolerance?: number | undefined;
     retryOnBlack?: boolean | undefined;
 }>;
@@ -1837,6 +1915,29 @@ 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;
@@ -1845,6 +1946,13 @@ declare const podcastInputSchema: z.ZodObject<{
     voices: Record<string, string>;
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     purpose?: string | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
     override?: {
         provider?: string | undefined;
         model?: string | undefined;
@@ -1865,6 +1973,13 @@ declare const podcastInputSchema: z.ZodObject<{
     voices: Record<string, string>;
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     purpose?: string | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
     override?: {
         provider?: string | undefined;
         model?: string | undefined;
@@ -1932,6 +2047,8 @@ declare const ttsInputSchema: z.ZodObject<{
         ipa?: string | undefined;
         lang?: string | undefined;
     }>, "many">>;
+    /** F055 — per-word timings (Azure only). Changes the route to batch synthesis. */
+    wordTimings: z.ZodOptional<z.ZodBoolean>;
     /** F037: voice to use if `voice` is one we know the provider has retired. Without
      *  it a retired voice throws VoiceUnavailableError rather than reaching the API.
      *
@@ -1951,6 +2068,14 @@ declare const ttsInputSchema: z.ZodObject<{
     voice: string;
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     purpose?: string | undefined;
+    wordTimings?: boolean | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
     override?: {
         provider?: string | undefined;
         model?: string | undefined;
@@ -1962,15 +2087,8 @@ declare const ttsInputSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     })[] | undefined;
     labels?: Record<string, string> | undefined;
-    format?: string | undefined;
     lang?: string | undefined;
-    pronunciations?: {
-        word: string;
-        matchInCompounds?: boolean | undefined;
-        alias?: string | undefined;
-        ipa?: string | undefined;
-        lang?: string | undefined;
-    }[] | undefined;
+    format?: string | undefined;
     voiceFallback?: string | undefined;
     rate?: number | undefined;
 }, {
@@ -1978,6 +2096,14 @@ declare const ttsInputSchema: z.ZodObject<{
     voice: string;
     tier?: "fast" | "smart" | "powerful" | "cheap" | "vision" | "video" | "embedding" | undefined;
     purpose?: string | undefined;
+    wordTimings?: boolean | undefined;
+    pronunciations?: {
+        word: string;
+        matchInCompounds?: boolean | undefined;
+        alias?: string | undefined;
+        ipa?: string | undefined;
+        lang?: string | undefined;
+    }[] | undefined;
     override?: {
         provider?: string | undefined;
         model?: string | undefined;
@@ -1989,15 +2115,8 @@ declare const ttsInputSchema: z.ZodObject<{
         transport: "http" | "subprocess";
     })[] | undefined;
     labels?: Record<string, string> | undefined;
-    format?: string | undefined;
     lang?: string | undefined;
-    pronunciations?: {
-        word: string;
-        matchInCompounds?: boolean | undefined;
-        alias?: string | undefined;
-        ipa?: string | undefined;
-        lang?: string | undefined;
-    }[] | undefined;
+    format?: string | undefined;
     voiceFallback?: string | undefined;
     rate?: number | undefined;
 }>;
@@ -2271,6 +2390,11 @@ declare function azureAdapter(config?: {
     resource?: string;
     /** Fast-transcription api-version (overrides the GA default). */
     sttApiVersion?: string;
+    /** F055 — how long to wait for a batch synthesis job. Default 180s (Microsoft's own
+     *  95th percentile is 120s). */
+    batchTimeoutMs?: number;
+    /** F055 — poll interval for batch synthesis. Default 3s. */
+    batchPollMs?: number;
     /** phraseList biasing weight (0–2) applied when a call passes `phrases`. Default 1.5. */
     sttBiasingWeight?: number;
 }): ProviderAdapter;
@@ -2390,8 +2514,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.43.0";
-declare const SDK_TAG: "@broberg/ai-sdk@0.43.0";
+declare const VERSION: "0.45.0";
+declare const SDK_TAG: "@broberg/ai-sdk@0.45.0";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.
@@ -2802,4 +2926,4 @@ interface StreamTransportRequest extends TransportRequest {
  */
 declare function streamTransport(req: StreamTransportRequest): AsyncIterable<string>;
 
-export { AZURE_DANISH_VOICES, AZURE_DANISH_VOICE_LIST, type AiClient, type AiConfig, type AzureVoiceInfo, type BatchJob, type BatchRequestItem, type BatchResultItem, type BflAdapterConfig, type BflCredits, type BudgetConfig, BudgetExceededError, BudgetGuard, type BudgetStore, type CallOptions, type Capability, type ChatInput, type ChatRequest, type ChatResult, type ChatStreamEvent, type CheckVoiceOptions, type ClassifyInput, type ClassifyResult, type ContentPart, type Contracts, type CostQuery, type CostSink, type CostSummary, type CostSummaryQuery, type CostTimeseriesQuery, DEFAULT_TIER_MAP, type DesignInput, type DesignResult, type DialogueRequest, type DialogueTurn, type DiscordSinkConfig, ELEVENLABS_DANISH_VOICES, type EmbeddingInput, type EmbeddingRequest, type EmbeddingResult, type ExtractInput, type ExtractResult, type FalAdapterConfig, type HttpResponse, type ImageInput, type ImageRequest, type ImageResult, type LoraWeight, type Message, type MockupInput, type MockupResult, type ModerationInput, type ModerationItem, type ModerationRequest, type ModerationResult, type OcrInput, type OcrPage, type OcrRequest, type OcrResult, type OpenAICompatibleConfig, type PodcastInput, type PodcastResult, type PricingEntry, type ProviderAdapter, type RefreshOptions, type RefreshResult, type Region, type RerankInput, type RerankResult, type Role, SDK_TAG, type SqliteBudgetStoreConfig, type SqliteSinkConfig, StreamHttpError, type SubprocessResponse, type Tier, type TierSpec, type Tool, type ToolCall, type TrainStyleInput, type TrainStyleRequest, type TrainStyleResult, type TranscribeInput, type TranscribeRequest, type TranscribeResult, type TranslateInput, type TranslateResult, type Transport, type TransportRequest, type TransportResponse, type TtsInput, type TtsRequest, type UpmetricsCostClientConfig, UpmetricsCostError, type UpmetricsCostRow, type UpmetricsCostSummary, type UpmetricsCostTimeseries, type UpmetricsSinkConfig, type Usage, VERSION, type VideoInput, type VisionInput, type VoiceInfo, type VoiceProvider, type VoiceResolveResult, type VoiceStatus, VoiceUnavailableError, aiConfigSchema, anthropicAdapter, anthropicApiAdapter, anthropicSubprocessAdapter, azureAdapter, bflAdapter, bflCredits, chatInputSchema, checkVoice, classifyRegionName, computeCost, createAI, deepinfraAdapter, deeplAdapter, deepseekAdapter, defaultProviders, discordSink, elevenlabsAdapter, embeddingInputSchema, falAdapter, falStubAdapter, freshUsage, fromProviderToolCall, geminiAdapter, getCostSummary, getPrice, httpTransport, imageInputSchema, listAzureDanishVoices, listVoices, makeContracts, makeOpenAICompatibleAdapter, messageSchema, mistralAdapter, mistralStubAdapter, multiSink, noopSink, openaiAdapter, openaiStubAdapter, openrouterAdapter, parseClaudeCliJson, parseJsonLoose, refreshAvailability, regionOfHost, regionOfProvider, requestyAdapter, resetRefreshState, resetRegistry, resolveAzureVoice, resolveTier, resolveVoice, sqliteBudgetStore, sqliteSink, streamTransport, stubProviders, subprocessTransport, tierSpecSchema, toProviderTools, toolSchema, translateInputSchema, upmetricsCostClient, upmetricsSink, usdFromMicro, vertexAdapter, visionInputSchema };
+export { AZURE_DANISH_VOICES, AZURE_DANISH_VOICE_LIST, type AiClient, type AiConfig, type AlignedWordTimings, type AzureVoiceInfo, type AzureWordBoundary, type BatchJob, type BatchRequestItem, type BatchResultItem, type BflAdapterConfig, type BflCredits, type BudgetConfig, BudgetExceededError, BudgetGuard, type BudgetStore, type CallOptions, type Capability, type ChatInput, type ChatRequest, type ChatResult, type ChatStreamEvent, type CheckVoiceOptions, type ClassifyInput, type ClassifyResult, type ContentPart, type Contracts, type CostQuery, type CostSink, type CostSummary, type CostSummaryQuery, type CostTimeseriesQuery, DEFAULT_TIER_MAP, type DesignInput, type DesignResult, type DialogueRequest, type DialogueTurn, type DiscordSinkConfig, ELEVENLABS_DANISH_VOICES, type EmbeddingInput, type EmbeddingRequest, type EmbeddingResult, type ExtractInput, type ExtractResult, type FalAdapterConfig, type HttpResponse, type ImageInput, type ImageRequest, type ImageResult, type LoraWeight, type Message, type MockupInput, type MockupResult, type ModerationInput, type ModerationItem, type ModerationRequest, type ModerationResult, type OcrInput, type OcrPage, type OcrRequest, type OcrResult, type OpenAICompatibleConfig, type PodcastInput, type PodcastResult, type PricingEntry, type ProviderAdapter, type RefreshOptions, type RefreshResult, type Region, type RerankInput, type RerankResult, type Role, SDK_TAG, type SqliteBudgetStoreConfig, type SqliteSinkConfig, StreamHttpError, type SubprocessResponse, type Tier, type TierSpec, type Tool, type ToolCall, type TrainStyleInput, type TrainStyleRequest, type TrainStyleResult, type TranscribeInput, type TranscribeRequest, type TranscribeResult, type TranslateInput, type TranslateResult, type Transport, type TransportRequest, type TransportResponse, type TtsInput, type TtsRequest, type UpmetricsCostClientConfig, UpmetricsCostError, type UpmetricsCostRow, type UpmetricsCostSummary, type UpmetricsCostTimeseries, type UpmetricsSinkConfig, type Usage, VERSION, type VideoInput, type VisionInput, type VoiceInfo, type VoiceProvider, type VoiceResolveResult, type VoiceStatus, VoiceUnavailableError, type WordTiming, aiConfigSchema, alignWordTimings, anthropicAdapter, anthropicApiAdapter, anthropicSubprocessAdapter, azureAdapter, bflAdapter, bflCredits, chatInputSchema, checkVoice, classifyRegionName, computeCost, createAI, deepinfraAdapter, deeplAdapter, deepseekAdapter, defaultProviders, discordSink, elevenlabsAdapter, embeddingInputSchema, falAdapter, falStubAdapter, freshUsage, fromProviderToolCall, geminiAdapter, getCostSummary, getPrice, httpTransport, imageInputSchema, listAzureDanishVoices, listVoices, makeContracts, makeOpenAICompatibleAdapter, messageSchema, mistralAdapter, mistralStubAdapter, multiSink, noopSink, openaiAdapter, openaiStubAdapter, openrouterAdapter, parseClaudeCliJson, parseJsonLoose, refreshAvailability, regionOfHost, regionOfProvider, requestyAdapter, resetRefreshState, resetRegistry, resolveAzureVoice, resolveTier, resolveVoice, sqliteBudgetStore, sqliteSink, streamTransport, stubProviders, subprocessTransport, tierSpecSchema, toProviderTools, toolSchema, translateInputSchema, upmetricsCostClient, upmetricsSink, usdFromMicro, vertexAdapter, visionInputSchema };

package/dist/index.js

@@ -1617,7 +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 } : {}
       })
     });
@@ -1652,6 +1658,124 @@ function elevenlabsAdapter(config = {}) {
   return { name: "elevenlabs", dialogue, tts, listVoices: listVoices2 };
 }
 
+// src/providers/word-timings.ts
+var PUNCTUATION = /* @__PURE__ */ new Set([".", ",", "!", "?", ":", ";", "\u2026", "\u2014", "\u2013", '"', "'", "(", ")"]);
+var fold = (s) => s.toLowerCase();
+function alignWordTimings(text, words, opts = {}) {
+  const out = [];
+  const unaligned = [];
+  const haystack = fold(text);
+  const aliases = (opts.pronunciations ?? []).filter((p) => typeof p.alias === "string" && p.alias.trim() !== "").map((p) => ({ word: p.word, parts: p.alias.trim().split(/\s+/).map(fold) })).sort((a, b) => b.parts.length - a.parts.length);
+  const spoken = words.filter((w) => !PUNCTUATION.has(w.Text.trim()));
+  let cursor = 0;
+  let i = 0;
+  while (i < spoken.length) {
+    let consumed = false;
+    for (const a of aliases) {
+      if (i + a.parts.length > spoken.length) continue;
+      const matches = a.parts.every((part, k) => fold(spoken[i + k].Text) === part);
+      if (!matches) continue;
+      const at2 = haystack.indexOf(fold(a.word), cursor);
+      if (at2 === -1) continue;
+      const span = { start: at2, end: at2 + a.word.length };
+      for (let k = 0; k < a.parts.length; k++) {
+        const w2 = spoken[i + k];
+        out.push({
+          text: w2.Text,
+          startMs: w2.AudioOffset,
+          endMs: w2.AudioOffset + w2.Duration,
+          sourceStart: span.start,
+          sourceEnd: span.end
+        });
+      }
+      cursor = span.end;
+      i += a.parts.length;
+      consumed = true;
+      break;
+    }
+    if (consumed) continue;
+    const w = spoken[i];
+    const at = haystack.indexOf(fold(w.Text), cursor);
+    if (at === -1) {
+      unaligned.push(w.Text);
+    } else {
+      out.push({
+        text: w.Text,
+        startMs: w.AudioOffset,
+        endMs: w.AudioOffset + w.Duration,
+        sourceStart: at,
+        sourceEnd: at + w.Text.length
+      });
+      cursor = at + w.Text.length;
+    }
+    i++;
+  }
+  return { words: out, unaligned };
+}
+
+// src/providers/zip.ts
+import { inflateRawSync } from "zlib";
+var EOCD_SIG = 101010256;
+var CEN_SIG = 33639248;
+function readZipEntry(zip, name) {
+  const view = new DataView(zip.buffer, zip.byteOffset, zip.byteLength);
+  let eocd = -1;
+  for (let i = zip.length - 22; i >= 0; i--) {
+    if (view.getUint32(i, true) === EOCD_SIG) {
+      eocd = i;
+      break;
+    }
+  }
+  if (eocd === -1) throw new Error("readZipEntry: not a ZIP archive (no end-of-central-directory record)");
+  const entryCount = view.getUint16(eocd + 10, true);
+  let p = view.getUint32(eocd + 16, true);
+  for (let n = 0; n < entryCount; n++) {
+    if (view.getUint32(p, true) !== CEN_SIG) {
+      throw new Error(`readZipEntry: corrupt central directory at entry ${n}`);
+    }
+    const method = view.getUint16(p + 10, true);
+    const compressedSize = view.getUint32(p + 20, true);
+    const nameLen = view.getUint16(p + 28, true);
+    const extraLen = view.getUint16(p + 30, true);
+    const commentLen = view.getUint16(p + 32, true);
+    const localOffset = view.getUint32(p + 42, true);
+    const entryName = new TextDecoder().decode(zip.subarray(p + 46, p + 46 + nameLen));
+    if (entryName === name) {
+      const lNameLen = view.getUint16(localOffset + 26, true);
+      const lExtraLen = view.getUint16(localOffset + 28, true);
+      const start = localOffset + 30 + lNameLen + lExtraLen;
+      const raw = zip.subarray(start, start + compressedSize);
+      if (method === 0) return new Uint8Array(raw);
+      if (method === 8) return new Uint8Array(inflateRawSync(raw));
+      throw new Error(`readZipEntry: "${name}" uses unsupported compression method ${method}`);
+    }
+    p += 46 + nameLen + extraLen + commentLen;
+  }
+  return void 0;
+}
+function listZipEntries(zip) {
+  const view = new DataView(zip.buffer, zip.byteOffset, zip.byteLength);
+  let eocd = -1;
+  for (let i = zip.length - 22; i >= 0; i--) {
+    if (view.getUint32(i, true) === EOCD_SIG) {
+      eocd = i;
+      break;
+    }
+  }
+  if (eocd === -1) throw new Error("listZipEntries: not a ZIP archive (no end-of-central-directory record)");
+  const entryCount = view.getUint16(eocd + 10, true);
+  let p = view.getUint32(eocd + 16, true);
+  const names = [];
+  for (let n = 0; n < entryCount; n++) {
+    const nameLen = view.getUint16(p + 28, true);
+    const extraLen = view.getUint16(p + 30, true);
+    const commentLen = view.getUint16(p + 32, true);
+    names.push(new TextDecoder().decode(zip.subarray(p + 46, p + 46 + nameLen)));
+    p += 46 + nameLen + extraLen + commentLen;
+  }
+  return names;
+}
+
 // src/providers/azure.ts
 var DEFAULT_STT_API_VERSION = "2025-10-15";
 var DEFAULT_REGION = "westeurope";
@@ -1733,10 +1857,9 @@ function azureAdapter(config = {}) {
     usage.costUsd = chars / 1e3 * (config.pricePer1kChars ?? getMediaPrice("azure", "tts")?.usd ?? 0);
     return usage;
   }
-  async function tts(req) {
+  function buildSsml(req) {
     const voice = resolveAzureVoice(req.voiceId);
     const lang = req.lang ?? localeOf(voice);
-    const format = req.format ?? DEFAULT_FORMAT;
     assertPronunciations(req.pronunciations, "azure");
     const escaped = applyPronunciations(
       xmlEscape(req.text),
@@ -1746,7 +1869,11 @@ function azureAdapter(config = {}) {
     );
     const effRate = req.rate ?? AZURE_DANISH_VOICE_LIST.find((v) => v.voiceId === voice)?.defaultRate;
     const inner = effRate != null && effRate !== 1 ? `<prosody rate='${effRate}'>${escaped}</prosody>` : escaped;
-    const ssml = `<speak version='1.0' xml:lang='${lang}'><voice name='${voice}'>${inner}</voice></speak>`;
+    return `<speak version='1.0' xml:lang='${lang}'><voice name='${voice}'>${inner}</voice></speak>`;
+  }
+  async function tts(req) {
+    if (req.wordTimings) return ttsBatch(req);
+    const format = req.format ?? DEFAULT_FORMAT;
     const res = await fetchImpl(
       `https://${region()}.tts.speech.microsoft.com/cognitiveservices/v1`,
       {
@@ -1756,7 +1883,7 @@ function azureAdapter(config = {}) {
           "Content-Type": "application/ssml+xml",
           "X-Microsoft-OutputFormat": format
         },
-        body: ssml
+        body: buildSsml(req)
       }
     );
     if (!res.ok) {
@@ -1766,6 +1893,82 @@ function azureAdapter(config = {}) {
     const audio = new Uint8Array(await res.arrayBuffer());
     return { audio, mimeType: "audio/mpeg", usage: priceFor(req.text.length, req.spec.model) };
   }
+  async function ttsBatch(req) {
+    const host = sttBaseUrl();
+    const api = "api-version=2024-04-01";
+    const id = `wt-${Date.now()}-${Math.random().toString(36).slice(2, 10)}`;
+    const headers = { "Ocp-Apim-Subscription-Key": key(), "Content-Type": "application/json" };
+    const put = await fetchImpl(`${host}/texttospeech/batchsyntheses/${id}?${api}`, {
+      method: "PUT",
+      headers,
+      body: JSON.stringify({
+        inputKind: "SSML",
+        inputs: [{ content: buildSsml(req) }],
+        properties: {
+          wordBoundaryEnabled: true,
+          // ONE audio file and ONE word list for the whole text. Without it a chunked
+          // input yields per-chunk offsets that somebody has to re-base — the kind of
+          // arithmetic that looks right and is 40 ms wrong by the end.
+          concatenateResult: true,
+          ...req.format ? { outputFormat: req.format } : {}
+        }
+      })
+    });
+    if (!put.ok) {
+      const body = await put.text().catch(() => "");
+      throw new Error(`azure batch synthesis submit ${put.status}: ${body.slice(0, 300)}`);
+    }
+    const deadline = Date.now() + (config.batchTimeoutMs ?? 18e4);
+    let resultUrl = "";
+    for (; ; ) {
+      const get = await fetchImpl(`${host}/texttospeech/batchsyntheses/${id}?${api}`, {
+        headers: { "Ocp-Apim-Subscription-Key": key() }
+      });
+      if (!get.ok) {
+        const body = await get.text().catch(() => "");
+        throw new Error(`azure batch synthesis poll ${get.status}: ${body.slice(0, 300)}`);
+      }
+      const job = await get.json();
+      if (job.status === "Succeeded") {
+        resultUrl = job.outputs?.result ?? "";
+        break;
+      }
+      if (job.status === "Failed") throw new Error(`azure batch synthesis ${id} failed`);
+      if (Date.now() > deadline) {
+        throw new Error(
+          `azure batch synthesis ${id} still "${job.status}" after ${Math.round((config.batchTimeoutMs ?? 18e4) / 1e3)}s`
+        );
+      }
+      await new Promise((r) => setTimeout(r, config.batchPollMs ?? 3e3));
+    }
+    if (!resultUrl) throw new Error(`azure batch synthesis ${id} succeeded without a result URL`);
+    const zipRes = await fetchImpl(resultUrl, { headers: { "Ocp-Apim-Subscription-Key": key() } });
+    if (!zipRes.ok) {
+      throw new Error(`azure batch synthesis results ${zipRes.status}`);
+    }
+    const zip = new Uint8Array(await zipRes.arrayBuffer());
+    const audio = readZipEntry(zip, "0001.wav");
+    if (!audio) {
+      throw new Error(`azure batch synthesis: no 0001.wav in results; archive holds ${listZipEntries(zip).join(", ")}`);
+    }
+    const wordsRaw = readZipEntry(zip, "0001.word.json");
+    if (!wordsRaw) {
+      throw new Error(
+        `azure batch synthesis: wordTimings was requested but the archive has no 0001.word.json; it holds ${listZipEntries(zip).join(", ")}`
+      );
+    }
+    const boundaries = JSON.parse(new TextDecoder().decode(wordsRaw));
+    return {
+      audio,
+      // Batch defaults to riff PCM, not mp3 — saying audio/mpeg here would be a lie the
+      // browser would act on.
+      mimeType: req.format?.includes("mp3") ? "audio/mpeg" : "audio/wav",
+      // Aligned against the ORIGINAL text, with the dictionary, so the offsets index the
+      // manuscript rather than the SSML we sent.
+      wordTimings: alignWordTimings(req.text, boundaries, { pronunciations: req.pronunciations }),
+      usage: priceFor(req.text.length, req.spec.model)
+    };
+  }
   async function transcribe(req) {
     const locale = toAzureLocale(req.language);
     const definition = { locales: [locale] };
@@ -3055,12 +3258,6 @@ var moderationInputSchema = z.object({
   input: z.union([z.string(), z.array(z.string())]),
   ...callOptions
 });
-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(),
-  ...callOptions
-});
 var pronunciationSchema = z.object({
   word: z.string(),
   alias: z.string().optional(),
@@ -3069,12 +3266,24 @@ var pronunciationSchema = z.object({
   /** 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 ttsInputSchema = z.object({
   text: z.string(),
   voice: z.string(),
   /** F051 — see TtsRequest.pronunciations. The alias/ipa exclusivity is enforced in
    *  the adapter, not here: the message must name the provider that cannot do it. */
   pronunciations: z.array(pronunciationSchema).optional(),
+  /** F055 — per-word timings (Azure only). Changes the route to batch synthesis. */
+  wordTimings: z.boolean().optional(),
   /** F037: voice to use if `voice` is one we know the provider has retired. Without
    *  it a retired voice throws VoiceUnavailableError rather than reaching the API.
    *
@@ -3116,8 +3325,8 @@ var aiConfigSchema = z.object({
 });
 
 // src/version.ts
-var VERSION = "0.43.0";
-var SDK_TAG = "@broberg/ai-sdk@0.43.0";
+var VERSION = "0.45.0";
+var SDK_TAG = "@broberg/ai-sdk@0.45.0";
 
 // src/cost/sinks/upmetrics.ts
 function upmetricsSink(config) {
@@ -3646,7 +3855,7 @@ 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 });
         }
       });
     },
@@ -3672,6 +3881,7 @@ function createAI(config = {}) {
             format: input.format,
             rate: input.rate,
             pronunciations: input.pronunciations,
+            wordTimings: input.wordTimings,
             spec
           });
         }
@@ -4140,6 +4350,7 @@ export {
   VERSION,
   VoiceUnavailableError,
   aiConfigSchema,
+  alignWordTimings,
   anthropicAdapter,
   anthropicApiAdapter,
   anthropicSubprocessAdapter,