@ssml-builder-js/azure-tts-client 2.12.0 → 2.14.0

package/src/synthesis.ts

@@ -1,7 +1,182 @@
 import * as SpeechSDK from "microsoft-cognitiveservices-speech-sdk";
-import { createSpeechSdkError } from "./errors.ts";
+import { createSpeechSdkError, UnsupportedMergeFormatError } from "./errors.ts";
 import { createSpeechConfig } from "./speechConfig.ts";
-import type { SsmlSynthesisResult, TtsConfig } from "./types.ts";
+import type { SsmlSynthesisChunk, SsmlSynthesisResult, SynthesisProgressEvent, TtsConfig } from "./types.ts";
+
+export type MergeAudioFormat = "wav" | "mp3" | "raw";
+
+function ascii(bytes: Uint8Array, offset: number, value: string): boolean {
+  return [...value].every((character, index) => bytes[offset + index] === character.charCodeAt(0));
+}
+
+function readUint32(bytes: Uint8Array, offset: number): number {
+  return new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength).getUint32(offset, true);
+}
+
+interface RiffChunk {
+  id: string;
+  data: Uint8Array;
+}
+
+interface ParsedWav {
+  chunks: RiffChunk[];
+  data: Uint8Array;
+  format: Uint8Array;
+}
+
+function parseWav(buffer: ArrayBuffer): ParsedWav {
+  const bytes = new Uint8Array(buffer);
+  if (bytes.byteLength < 12 || !ascii(bytes, 0, "RIFF") || !ascii(bytes, 8, "WAVE")) {
+    throw new Error("Invalid WAV/RIFF audio buffer.");
+  }
+  const chunks: RiffChunk[] = [];
+  const dataParts: Uint8Array[] = [];
+  let format: Uint8Array | undefined;
+  let offset = 12;
+  while (offset < bytes.byteLength) {
+    if (offset + 8 > bytes.byteLength) throw new Error("Invalid WAV chunk header.");
+    const id = String.fromCharCode(...bytes.slice(offset, offset + 4));
+    const size = readUint32(bytes, offset + 4);
+    const dataStart = offset + 8;
+    const dataEnd = dataStart + size;
+    if (dataEnd > bytes.byteLength) throw new Error(`WAV chunk "${id}" exceeds the audio buffer.`);
+    const data = bytes.slice(dataStart, dataEnd);
+    chunks.push({ id, data });
+    if (id === "fmt ") format ??= data;
+    if (id === "data") dataParts.push(data);
+    offset = dataEnd + (size & 1);
+    if (offset > bytes.byteLength) throw new Error("Invalid WAV chunk padding.");
+  }
+  if (!format || dataParts.length === 0) throw new Error("WAV audio must contain fmt and data chunks.");
+  const dataLength = dataParts.reduce((total, part) => total + part.byteLength, 0);
+  const data = new Uint8Array(dataLength);
+  let dataOffset = 0;
+  for (const part of dataParts) {
+    data.set(part, dataOffset);
+    dataOffset += part.byteLength;
+  }
+  return { chunks, data, format };
+}
+
+function writeUint32(target: Uint8Array, offset: number, value: number): void {
+  new DataView(target.buffer).setUint32(offset, value, true);
+}
+
+function writeChunk(target: Uint8Array, offset: number, id: string, data: Uint8Array): number {
+  for (let index = 0; index < 4; index += 1) target[offset + index] = id.charCodeAt(index) ?? 0;
+  writeUint32(target, offset + 4, data.byteLength);
+  target.set(data, offset + 8);
+  const end = offset + 8 + data.byteLength;
+  if (data.byteLength & 1) target[end] = 0;
+  return end + (data.byteLength & 1);
+}
+
+function mergeWavBuffers(buffers: readonly ArrayBuffer[]): ArrayBuffer {
+  if (buffers.length === 0) return new ArrayBuffer(0);
+  const parsed = buffers.map(parseWav);
+  const first = parsed[0];
+  if (!first) throw new Error("At least one WAV buffer is required.");
+  if (
+    parsed.some(
+      (item) => item.format.length !== first.format.length || item.format.some((value, i) => value !== first.format[i]),
+    )
+  )
+    throw new Error("WAV buffers have incompatible fmt chunks.");
+  const dataLength = parsed.reduce((total, item) => total + item.data.byteLength, 0);
+  const nonDataLength = first.chunks.reduce(
+    (total, chunk) => (chunk.id === "data" ? total : total + 8 + chunk.data.byteLength + (chunk.data.byteLength & 1)),
+    0,
+  );
+  const outputLength = 12 + nonDataLength + 8 + dataLength + (dataLength & 1);
+  if (outputLength - 8 > 0xffffffff) throw new RangeError("Merged WAV exceeds the RIFF format size limit.");
+  const output = new Uint8Array(outputLength);
+  output.set(Uint8Array.from([0x52, 0x49, 0x46, 0x46]), 0);
+  writeUint32(output, 4, outputLength - 8);
+  output.set(Uint8Array.from([0x57, 0x41, 0x56, 0x45]), 8);
+  let outputOffset = 12;
+  let dataWritten = false;
+  for (const chunk of first.chunks) {
+    if (chunk.id === "data") {
+      if (dataWritten) continue;
+      const data = new Uint8Array(dataLength);
+      let dataOffset = 0;
+      for (const item of parsed) {
+        data.set(item.data, dataOffset);
+        dataOffset += item.data.byteLength;
+      }
+      outputOffset = writeChunk(output, outputOffset, "data", data);
+      dataWritten = true;
+    } else {
+      outputOffset = writeChunk(output, outputOffset, chunk.id, chunk.data);
+    }
+  }
+  if (!dataWritten) throw new Error("WAV audio must contain a data chunk.");
+  return output.buffer;
+}
+
+function skipId3v2(bytes: Uint8Array): number {
+  if (!ascii(bytes, 0, "ID3") || bytes.byteLength < 10) return 0;
+  const size = [bytes[6], bytes[7], bytes[8], bytes[9]].reduce((total, value) => (total << 7) | (value & 0x7f), 0);
+  const hasFooter = (bytes[5] & 0x10) !== 0;
+  return Math.min(bytes.byteLength, 10 + size + (hasFooter ? 10 : 0));
+}
+
+function stripMp3Tags(buffer: ArrayBuffer): Uint8Array {
+  const bytes = new Uint8Array(buffer);
+  const start = skipId3v2(bytes);
+  const end =
+    bytes.byteLength >= 128 && ascii(bytes, bytes.byteLength - 128, "TAG") ? bytes.byteLength - 128 : bytes.byteLength;
+  return bytes.slice(Math.min(start, end), end);
+}
+
+function isMp3Format(format: string): boolean {
+  return /(?:mp3|mpeg)/i.test(format);
+}
+
+function isWavFormat(format: string): boolean {
+  return /(?:wav|wave|riff)/i.test(format);
+}
+
+function isRawFormat(format: string): boolean {
+  return /^raw(?:-|$)/i.test(format);
+}
+
+/** Returns whether the named output format can be safely concatenated without re-multiplexing. */
+export function resolveMergeAudioFormat(format: string): MergeAudioFormat | undefined {
+  if (isWavFormat(format)) return "wav";
+  if (isMp3Format(format)) return "mp3";
+  if (isRawFormat(format)) return "raw";
+  return undefined;
+}
+
+export function canMergeAudioFormat(format: string): boolean {
+  return resolveMergeAudioFormat(format) !== undefined;
+}
+
+/** Merges audio buffers while preserving the invariants of supported containers. */
+export function mergeAudioBuffers(buffers: readonly ArrayBuffer[], format: string): ArrayBuffer {
+  if (isWavFormat(format)) return mergeWavBuffers(buffers);
+  if (isMp3Format(format)) {
+    const parts = buffers.map(stripMp3Tags);
+    const output = new Uint8Array(parts.reduce((total, part) => total + part.byteLength, 0));
+    let offset = 0;
+    for (const part of parts) {
+      output.set(part, offset);
+      offset += part.byteLength;
+    }
+    return output.buffer;
+  }
+  if (isRawFormat(format)) {
+    const output = new Uint8Array(buffers.reduce((total, buffer) => total + buffer.byteLength, 0));
+    let offset = 0;
+    for (const buffer of buffers) {
+      output.set(new Uint8Array(buffer), offset);
+      offset += buffer.byteLength;
+    }
+    return output.buffer;
+  }
+  throw new UnsupportedMergeFormatError(format);
+}
 
 function closeSpeechResources(speechConfig: SpeechSDK.SpeechConfig, synthesizer: SpeechSDK.SpeechSynthesizer): void {
   try {
@@ -80,12 +255,28 @@ export async function synthesizeSsml(ssml: string, config: TtsConfig): Promise<S
         ...(bookmarks ?? []).map((bookmark) => bookmark.audioOffsetMs),
       );
       const durationMs = result.audioDuration ? ticksToMilliseconds(result.audioDuration) : eventDurationMs;
+      const requestId = (result as SpeechSDK.SpeechSynthesisResult & { resultId?: string }).resultId;
+      const addSourceMetadata = <T extends { audioOffsetMs: number }>(event: T): T => ({
+        ...event,
+        ...(config.sourceTextRange ? { textRange: { ...config.sourceTextRange } } : {}),
+        ...(config.sourceTextRange ? { originalTextRange: { ...config.sourceTextRange } } : {}),
+        ...(config.chunkIndex !== undefined ? { chunkIndex: config.chunkIndex } : {}),
+        ...(config.sourceNodePath ? { sourceNodePath: [...config.sourceNodePath] } : {}),
+        ...(requestId ? { requestId } : {}),
+      });
+      const sourceBoundaries = boundaries.map((boundary) => addSourceMetadata(boundary));
+      const sourceVisemes = visemes.map((viseme) => addSourceMetadata(viseme));
+      const sourceBookmarks = bookmarks.map((bookmark) => addSourceMetadata(bookmark));
       resolve({
         audioData: result.audioData,
         durationMs,
-        ...(boundaries.length > 0 ? { boundaries, wordBoundary: boundaries, wordBoundaries: boundaries } : {}),
-        ...(visemes.length > 0 ? { visemes } : {}),
-        ...(bookmarks.length > 0 ? { bookmarks } : {}),
+        ...(config.sourceTextRange ? { textRange: { ...config.sourceTextRange } } : {}),
+        ...(requestId ? { requestId } : {}),
+        ...(sourceBoundaries.length > 0
+          ? { boundaries: sourceBoundaries, wordBoundary: sourceBoundaries, wordBoundaries: sourceBoundaries }
+          : {}),
+        ...(sourceVisemes.length > 0 ? { visemes: sourceVisemes } : {}),
+        ...(sourceBookmarks.length > 0 ? { bookmarks: sourceBookmarks } : {}),
       });
     };
 
@@ -107,6 +298,159 @@ export async function synthesizeSsml(ssml: string, config: TtsConfig): Promise<S
   });
 }
 
+/** Synthesizes chunks sequentially, annotates synchronization events, and merges the results. */
+export async function synthesizeSsmlChunks(
+  chunks: readonly (SsmlSynthesisChunk | string)[],
+  config: TtsConfig,
+): Promise<SsmlSynthesisResult> {
+  const results: SsmlSynthesisResult[] = [];
+  const totalChunks = chunks.length;
+  const report = (event: SynthesisProgressEvent): void => config.onProgress?.(event);
+  for (const [index, chunk] of chunks.entries()) {
+    const input = typeof chunk === "string" ? { ssml: chunk } : chunk;
+    report({
+      currentChunk: index,
+      totalChunks,
+      percent: totalChunks === 0 ? 100 : Math.round((index / totalChunks) * 100),
+      chunkIndex: index,
+      originalTextRange: input.originalTextRange,
+      status: "pending",
+      durationMs: 0,
+    });
+  }
+  for (const [index, chunk] of chunks.entries()) {
+    const input = typeof chunk === "string" ? { ssml: chunk } : chunk;
+    report({
+      currentChunk: index,
+      totalChunks,
+      percent: totalChunks === 0 ? 100 : Math.round((index / totalChunks) * 100),
+      chunkIndex: index,
+      originalTextRange: input.originalTextRange,
+      status: "synthesizing",
+      durationMs: 0,
+    });
+    const startedAt = Date.now();
+    try {
+      const result = await synthesizeSsml(input.ssml, {
+        ...config,
+        ...(input.originalTextRange ? { sourceTextRange: input.originalTextRange } : {}),
+        ...(input.sourceNodePath ? { sourceNodePath: input.sourceNodePath } : {}),
+        chunkIndex: index,
+        onProgress: undefined,
+      });
+      results.push(result);
+      report({
+        currentChunk: index + 1,
+        totalChunks,
+        percent: totalChunks === 0 ? 100 : Math.round(((index + 1) / totalChunks) * 100),
+        chunkIndex: index,
+        originalTextRange: input.originalTextRange,
+        status: "success",
+        durationMs: Date.now() - startedAt,
+      });
+    } catch (error) {
+      report({
+        currentChunk: index,
+        totalChunks,
+        percent: totalChunks === 0 ? 100 : Math.round((index / totalChunks) * 100),
+        chunkIndex: index,
+        originalTextRange: input.originalTextRange,
+        status: "failed",
+        durationMs: Date.now() - startedAt,
+        error,
+      });
+      throw error;
+    }
+  }
+  return mergeSynthesisResults(results, config.outputFormat ?? "audio-16khz-128kbitrate-mono-mp3");
+}
+
+/** Concatenates audio buffers and shifts all synchronization events by prior chunk durations. */
+export function mergeSynthesisResults(results: readonly SsmlSynthesisResult[], format?: string): SsmlSynthesisResult {
+  const audioData = format
+    ? new Uint8Array(
+        mergeAudioBuffers(
+          results.map((result) => result.audioData),
+          format,
+        ),
+      )
+    : new Uint8Array(results.reduce((total, result) => total + result.audioData.byteLength, 0));
+  if (!format) {
+    let offset = 0;
+    for (const result of results) {
+      audioData.set(new Uint8Array(result.audioData), offset);
+      offset += result.audioData.byteLength;
+    }
+  }
+  const boundaries: NonNullable<SsmlSynthesisResult["boundaries"]> = [];
+  const visemes: NonNullable<SsmlSynthesisResult["visemes"]> = [];
+  const bookmarks: NonNullable<SsmlSynthesisResult["bookmarks"]> = [];
+  let durationOffset = 0;
+
+  for (const result of results) {
+    const chunkBoundaries =
+      result.boundaries && result.boundaries.length > 0
+        ? result.boundaries
+        : (result.wordBoundary ?? result.wordBoundaries ?? []);
+    for (const boundary of chunkBoundaries) {
+      const textRange = boundary.textRange ?? result.textRange;
+      const originalTextRange = boundary.originalTextRange ?? textRange;
+      const requestId = boundary.requestId ?? result.requestId;
+      boundaries.push({
+        ...boundary,
+        audioOffsetMs: boundary.audioOffsetMs + durationOffset,
+        chunkAudioOffsetMs: boundary.chunkAudioOffsetMs ?? boundary.audioOffsetMs,
+        ...(boundary.chunkIndex === undefined ? { chunkIndex: results.indexOf(result) } : {}),
+        ...(boundary.sourceNodePath ? { sourceNodePath: [...boundary.sourceNodePath] } : {}),
+        ...(originalTextRange ? { originalTextRange: { ...originalTextRange } } : {}),
+        ...(textRange ? { textRange: { ...textRange } } : {}),
+        ...(requestId ? { requestId } : {}),
+      });
+    }
+    for (const viseme of result.visemes ?? []) {
+      const textRange = viseme.textRange ?? result.textRange;
+      const originalTextRange = viseme.originalTextRange ?? textRange;
+      const requestId = viseme.requestId ?? result.requestId;
+      visemes.push({
+        ...viseme,
+        audioOffsetMs: viseme.audioOffsetMs + durationOffset,
+        chunkAudioOffsetMs: viseme.chunkAudioOffsetMs ?? viseme.audioOffsetMs,
+        ...(viseme.chunkIndex === undefined ? { chunkIndex: results.indexOf(result) } : {}),
+        ...(viseme.sourceNodePath ? { sourceNodePath: [...viseme.sourceNodePath] } : {}),
+        ...(originalTextRange ? { originalTextRange: { ...originalTextRange } } : {}),
+        ...(textRange ? { textRange: { ...textRange } } : {}),
+        ...(requestId ? { requestId } : {}),
+      });
+    }
+    for (const bookmark of result.bookmarks ?? []) {
+      const textRange = bookmark.textRange ?? result.textRange;
+      const originalTextRange = bookmark.originalTextRange ?? textRange;
+      const requestId = bookmark.requestId ?? result.requestId;
+      bookmarks.push({
+        ...bookmark,
+        audioOffsetMs: bookmark.audioOffsetMs + durationOffset,
+        chunkAudioOffsetMs: bookmark.chunkAudioOffsetMs ?? bookmark.audioOffsetMs,
+        ...(bookmark.chunkIndex === undefined ? { chunkIndex: results.indexOf(result) } : {}),
+        ...(bookmark.sourceNodePath ? { sourceNodePath: [...bookmark.sourceNodePath] } : {}),
+        ...(originalTextRange ? { originalTextRange: { ...originalTextRange } } : {}),
+        ...(textRange ? { textRange: { ...textRange } } : {}),
+        ...(requestId ? { requestId } : {}),
+      });
+    }
+    durationOffset += Math.max(0, result.durationMs);
+  }
+
+  return {
+    audioData: audioData.buffer,
+    durationMs: durationOffset,
+    ...(boundaries.length > 0 ? { boundaries, wordBoundary: boundaries, wordBoundaries: boundaries } : {}),
+    ...(visemes.length > 0 ? { visemes } : {}),
+    ...(bookmarks.length > 0 ? { bookmarks } : {}),
+    ...(results.length === 1 && results[0]?.requestId ? { requestId: results[0].requestId } : {}),
+    ...(results.length === 1 && results[0]?.textRange ? { textRange: { ...results[0].textRange } } : {}),
+  };
+}
+
 /** Backward-compatible audio-only synthesis helper. */
 export async function synthesizeSpeech(ssml: string, config: TtsConfig): Promise<ArrayBuffer> {
   return (await synthesizeSsml(ssml, config)).audioData;

package/src/types.ts

@@ -1,3 +1,5 @@
+import type { SsmlTextRange } from "@ssml-builder-js/ssml-core";
+
 export interface TtsConfig {
   signal?: AbortSignal;
   timeoutMs?: number;
@@ -5,22 +7,53 @@ export interface TtsConfig {
   subscriptionKey: string;
   region: string;
   outputFormat?: string;
+  /** Original plain-text range represented by this synthesis request. */
+  sourceTextRange?: { start: number; end: number };
+  /** Reports chunk lifecycle events when using chunk synthesis. */
+  onProgress?: (event: SynthesisProgressEvent) => void;
+  /** Metadata used to map synchronization events back to the source document. */
+  chunkIndex?: number;
+  sourceNodePath?: string[];
 }
 
+export type SynthesisChunkStatus = "pending" | "synthesizing" | "success" | "failed";
+
 export interface SsmlSynthesisBoundary {
   text: string;
   audioOffsetMs: number;
   durationMs: number;
+  textRange?: { start: number; end: number };
+  /** Chunk that produced this event. */
+  chunkIndex?: number;
+  /** Path of the source SSML node, when available. */
+  sourceNodePath?: string[];
+  /** Original text range represented by this event. */
+  originalTextRange?: SsmlTextRange;
+  /** Audio offset within the originating chunk before merge. */
+  chunkAudioOffsetMs?: number;
+  requestId?: string;
 }
 
 export interface SsmlSynthesisViseme {
   visemeId: number;
   audioOffsetMs: number;
+  textRange?: { start: number; end: number };
+  chunkIndex?: number;
+  sourceNodePath?: string[];
+  originalTextRange?: SsmlTextRange;
+  chunkAudioOffsetMs?: number;
+  requestId?: string;
 }
 
 export interface SsmlSynthesisBookmark {
   name: string;
   audioOffsetMs: number;
+  textRange?: { start: number; end: number };
+  chunkIndex?: number;
+  sourceNodePath?: string[];
+  originalTextRange?: SsmlTextRange;
+  chunkAudioOffsetMs?: number;
+  requestId?: string;
 }
 
 /** Audio and Azure Speech synchronization events emitted for one SSML request. */
@@ -34,6 +67,32 @@ export interface SsmlSynthesisResult {
   wordBoundaries?: SsmlSynthesisBoundary[];
   visemes?: SsmlSynthesisViseme[];
   bookmarks?: SsmlSynthesisBookmark[];
+  /** Request identifier returned by Azure Speech, when available. */
+  requestId?: string;
+  /** Original plain-text range represented by the result. */
+  textRange?: { start: number; end: number };
+}
+
+export interface SsmlSynthesisChunk {
+  ssml: string;
+  originalTextRange?: { start: number; end: number };
+  sourceNodePath?: string[];
+}
+
+export interface SynthesizeChunksOptions {
+  onProgress?: (event: SynthesisProgressEvent) => void;
+}
+
+export interface SynthesisProgressEvent {
+  /** 1-based completed chunk count retained for backward compatibility. */
+  currentChunk: number;
+  totalChunks: number;
+  percent: number;
+  chunkIndex: number;
+  originalTextRange?: SsmlTextRange;
+  status: SynthesisChunkStatus;
+  durationMs: number;
+  error?: unknown;
 }
 
 export interface AzureTtsLogger {
@@ -51,4 +110,5 @@ export interface AzureTtsClientOptions {
   endpoint?: string;
   outputFormat?: string;
   logger?: AzureTtsLogger;
+  onProgress?: (event: SynthesisProgressEvent) => void;
 }

package/src/voiceCatalog.ts

@@ -10,6 +10,9 @@ export interface AzureVoiceCatalogVoice {
   locale: string;
   secondaryLocales?: readonly string[];
   styles?: readonly string[];
+  supportedTags?: readonly string[];
+  unsupportedTags?: readonly string[];
+  models?: readonly string[];
   regions: readonly string[];
   status?: "ga" | "preview" | "deprecated";
 }
@@ -33,6 +36,9 @@ interface AzureVoiceApiRecord {
   ShortName?: unknown;
   Status?: unknown;
   StyleList?: unknown;
+  SupportedTags?: unknown;
+  UnsupportedTags?: unknown;
+  Models?: unknown;
 }
 
 function stringValue(value: unknown): string | undefined {
@@ -92,6 +98,9 @@ export async function fetchAzureVoiceCatalog(options: FetchAzureVoiceCatalogOpti
       const secondaryLocales = stringList(record.SecondaryLocaleList);
       const styles = stringList(record.StyleList);
       const status = normalizeStatus(record.Status);
+      const supportedTags = stringList(record.SupportedTags);
+      const unsupportedTags = stringList(record.UnsupportedTags);
+      const models = stringList(record.Models);
       const merged: AzureVoiceCatalogVoice = {
         name: existing?.name ?? name,
         locale: existing?.locale ?? locale,
@@ -101,6 +110,12 @@ export async function fetchAzureVoiceCatalog(options: FetchAzureVoiceCatalogOpti
       if (mergedSecondaryLocales.length > 0) merged.secondaryLocales = mergedSecondaryLocales;
       const mergedStyles = [...new Set([...(existing?.styles ?? []), ...styles])];
       if (mergedStyles.length > 0) merged.styles = mergedStyles;
+      const mergedSupportedTags = [...new Set([...(existing?.supportedTags ?? []), ...supportedTags])];
+      if (mergedSupportedTags.length > 0) merged.supportedTags = mergedSupportedTags;
+      const mergedUnsupportedTags = [...new Set([...(existing?.unsupportedTags ?? []), ...unsupportedTags])];
+      if (mergedUnsupportedTags.length > 0) merged.unsupportedTags = mergedUnsupportedTags;
+      const mergedModels = [...new Set([...(existing?.models ?? []), ...models])];
+      if (mergedModels.length > 0) merged.models = mergedModels;
       if (status) merged.status = status;
       else if (existing?.status) merged.status = existing.status;
       voices.set(key, merged);

package/test/v213-pipeline.test.ts

@@ -0,0 +1,72 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { mergeSynthesisResults, synthesizeSsmlSafe } from "../src/index.ts";
+
+const audio = (values: number[]): ArrayBuffer => Uint8Array.from(values).buffer;
+
+test("mergeSynthesisResults concatenates audio and offsets synchronization events", () => {
+  const result = mergeSynthesisResults([
+    {
+      audioData: audio([1, 2]),
+      durationMs: 100,
+      boundaries: [{ text: "one", audioOffsetMs: 20, durationMs: 30, textRange: { start: 0, end: 3 }, requestId: "a" }],
+      visemes: [{ visemeId: 1, audioOffsetMs: 40 }],
+      bookmarks: [{ name: "first", audioOffsetMs: 50 }],
+    },
+    {
+      audioData: audio([3, 4, 5]),
+      durationMs: 250,
+      boundaries: [{ text: "two", audioOffsetMs: 10, durationMs: 20, textRange: { start: 3, end: 6 }, requestId: "b" }],
+      visemes: [{ visemeId: 2, audioOffsetMs: 15 }],
+      bookmarks: [{ name: "second", audioOffsetMs: 25 }],
+    },
+  ]);
+
+  assert.deepEqual([...new Uint8Array(result.audioData)], [1, 2, 3, 4, 5]);
+  assert.equal(result.durationMs, 350);
+  assert.deepEqual(
+    result.boundaries?.map(({ audioOffsetMs }) => audioOffsetMs),
+    [20, 110],
+  );
+  assert.deepEqual(
+    result.visemes?.map(({ audioOffsetMs }) => audioOffsetMs),
+    [40, 115],
+  );
+  assert.deepEqual(
+    result.bookmarks?.map(({ audioOffsetMs }) => audioOffsetMs),
+    [50, 125],
+  );
+  assert.deepEqual(result.boundaries?.[1]?.textRange, { start: 3, end: 6 });
+  assert.equal(result.boundaries?.[0]?.chunkIndex, 0);
+  assert.equal(result.boundaries?.[1]?.chunkIndex, 1);
+  assert.equal(result.boundaries?.[1]?.chunkAudioOffsetMs, 10);
+  assert.deepEqual(result.boundaries?.[1]?.originalTextRange, { start: 3, end: 6 });
+});
+
+test("synthesizeSsmlSafe blocks invalid SSML without calling the client", async () => {
+  let calls = 0;
+  const result = await synthesizeSsmlSafe(
+    {
+      synthesizeSsml: async () => {
+        calls += 1;
+        return { audioData: new ArrayBuffer(0), durationMs: 0 };
+      },
+    },
+    "<speak>",
+  );
+
+  assert.equal(result.ok, false);
+  assert.equal(result.status, "validation-error");
+  assert.equal(calls, 0);
+});
+
+test("synthesizeSsmlSafe returns a successful result for valid SSML", async () => {
+  const expected = { audioData: audio([1]), durationMs: 10 };
+  const result = await synthesizeSsmlSafe(
+    { synthesizeSsml: async () => expected },
+    '<speak version="1.0" xml:lang="en-US"><voice name="en-US-JennyNeural">Hello</voice></speak>',
+  );
+
+  assert.equal(result.ok, true);
+  if (result.ok) assert.equal(result.value, expected);
+});