@speech-sdk/core 0.31.0 → 0.33.0

package/dist/conversation/types.d.ts

@@ -27,6 +27,11 @@ export interface GenerateConversationOptions<V extends Voice = Voice, M extends
     readonly pronunciations?: PronunciationsInput;
     readonly providerOptions?: Record<string, unknown>;
     readonly speed?: number;
+    /**
+     * Also return `turns`: one audio clip per input turn, cut at the silence after each turn's last word. Requires
+     * `timestamps: true`. Throws `TurnSplitError` when turn boundaries can't be trusted.
+     */
+    readonly splitTurns?: boolean;
     readonly timestampProvider?: TimestampProvider;
     readonly timestamps?: boolean;
     readonly turns: readonly ConversationTurn<V>[];

package/dist/conversation/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/conversation/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACtD,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACtE,OAAO,KAAK,EAAE,aAAa,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AAClE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAElE,MAAM,WAAW,gBAAgB,CAAC,CAAC,SAAS,KAAK,GAAG,KAAK;IACvD,mDAAmD;IACnD,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,CAAC;IAC3C,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAEnD,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,0DAA0D;IAC1D,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC;CACnB;AAED,MAAM,WAAW,2BAA2B,CAC1C,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS;IAEb,QAAQ,CAAC,WAAW,CAAC,EAAE,WAAW,CAAC;IACnC,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC1C,iEAAiE;IACjE,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IACnB,QAAQ,CAAC,MAAM,CAAC,EAAE,WAAW,CAAC;IAC9B,QAAQ,CAAC,cAAc,CAAC,EAAE,mBAAmB,CAAC;IAC9C,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAEnD,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;IAC/C,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,SAAS,gBAAgB,CAAC,CAAC,CAAC,EAAE,CAAC;IAE/C,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;CAC9B"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/conversation/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACtD,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACtE,OAAO,KAAK,EAAE,aAAa,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AAClE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAElE,MAAM,WAAW,gBAAgB,CAAC,CAAC,SAAS,KAAK,GAAG,KAAK;IACvD,mDAAmD;IACnD,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,CAAC;IAC3C,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAEnD,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,0DAA0D;IAC1D,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC;CACnB;AAED,MAAM,WAAW,2BAA2B,CAC1C,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS;IAEb,QAAQ,CAAC,WAAW,CAAC,EAAE,WAAW,CAAC;IACnC,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC1C,iEAAiE;IACjE,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IACnB,QAAQ,CAAC,MAAM,CAAC,EAAE,WAAW,CAAC;IAC9B,QAAQ,CAAC,cAAc,CAAC,EAAE,mBAAmB,CAAC;IAC9C,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAEnD,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;;;OAGG;IACH,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,CAAC;IAC9B,QAAQ,CAAC,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;IAC/C,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,SAAS,gBAAgB,CAAC,CAAC,CAAC,EAAE,CAAC;IAE/C,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;CAC9B"}

package/dist/conversation/validate.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"validate.d.ts","sourceRoot":"","sources":["../../src/conversation/validate.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,gBAAgB,EAAE,2BAA2B,EAAE,MAAM,YAAY,CAAC;AAEhF,wBAAgB,aAAa,IAAI,CAAC,KAAK,EAAE,gBAAgB,CAAC,OAAO,CAAC,KAAK,MAAM,CAE5E;AAED,wBAAgB,yBAAyB,CACvC,OAAO,EAAE,2BAA2B,GACnC,IAAI,CA2BN"}
+{"version":3,"file":"validate.d.ts","sourceRoot":"","sources":["../../src/conversation/validate.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,gBAAgB,EAAE,2BAA2B,EAAE,MAAM,YAAY,CAAC;AAEhF,wBAAgB,aAAa,IAAI,CAAC,KAAK,EAAE,gBAAgB,CAAC,OAAO,CAAC,KAAK,MAAM,CAE5E;AAED,wBAAgB,yBAAyB,CACvC,OAAO,EAAE,2BAA2B,GACnC,IAAI,CAiCN"}

package/dist/conversation/validate.js

@@ -6,6 +6,9 @@ export function validateConversationInput(options) {
     if (options.turns.length === 0) {
         throw new ConversationInputError("generateConversation requires at least one turn.");
     }
+    if (options.splitTurns && options.timestamps !== true) {
+        throw new ConversationInputError("splitTurns requires timestamps: true — turn boundaries come from word timestamps.");
+    }
     // Model placement must be all-or-nothing — partial mix hides which model actually ran where.
     const hasTopLevel = options.model != null;
     for (let i = 0; i < options.turns.length; i++) {

package/dist/conversation/validate.js.map

@@ -1 +1 @@
-{"version":3,"file":"validate.js","sourceRoot":"","sources":["../../src/conversation/validate.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,sBAAsB,EAAE,MAAM,aAAa,CAAC;AAGrD,MAAM,UAAU,aAAa;IAC3B,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,EAAE,CAAC;AACjC,CAAC;AAED,MAAM,UAAU,yBAAyB,CACvC,OAAoC;IAEpC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC/B,MAAM,IAAI,sBAAsB,CAC9B,kDAAkD,CACnD,CAAC;IACJ,CAAC;IAED,6FAA6F;IAC7F,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC;IAE1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC9B,IAAI,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAClC,MAAM,IAAI,sBAAsB,CAAC,SAAS,CAAC,2BAA2B,CAAC,CAAC;QAC1E,CAAC;QACD,MAAM,YAAY,GAAG,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC;QACxC,IAAI,WAAW,IAAI,YAAY,EAAE,CAAC;YAChC,MAAM,IAAI,sBAAsB,CAC9B,SAAS,CAAC,kIAAkI,CAC7I,CAAC;QACJ,CAAC;QACD,IAAI,CAAC,CAAC,WAAW,IAAI,YAAY,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,sBAAsB,CAC9B,SAAS,CAAC,2HAA2H,CACtI,CAAC;QACJ,CAAC;IACH,CAAC;AACH,CAAC"}
+{"version":3,"file":"validate.js","sourceRoot":"","sources":["../../src/conversation/validate.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,sBAAsB,EAAE,MAAM,aAAa,CAAC;AAGrD,MAAM,UAAU,aAAa;IAC3B,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,EAAE,CAAC;AACjC,CAAC;AAED,MAAM,UAAU,yBAAyB,CACvC,OAAoC;IAEpC,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC/B,MAAM,IAAI,sBAAsB,CAC9B,kDAAkD,CACnD,CAAC;IACJ,CAAC;IAED,IAAI,OAAO,CAAC,UAAU,IAAI,OAAO,CAAC,UAAU,KAAK,IAAI,EAAE,CAAC;QACtD,MAAM,IAAI,sBAAsB,CAC9B,mFAAmF,CACpF,CAAC;IACJ,CAAC;IAED,6FAA6F;IAC7F,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC;IAE1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC9B,IAAI,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAClC,MAAM,IAAI,sBAAsB,CAAC,SAAS,CAAC,2BAA2B,CAAC,CAAC;QAC1E,CAAC;QACD,MAAM,YAAY,GAAG,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC;QACxC,IAAI,WAAW,IAAI,YAAY,EAAE,CAAC;YAChC,MAAM,IAAI,sBAAsB,CAC9B,SAAS,CAAC,kIAAkI,CAC7I,CAAC;QACJ,CAAC;QACD,IAAI,CAAC,CAAC,WAAW,IAAI,YAAY,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,sBAAsB,CAC9B,SAAS,CAAC,2HAA2H,CACtI,CAAC;QACJ,CAAC;IACH,CAAC;AACH,CAAC"}

package/dist/generate-conversation.d.ts

@@ -1,9 +1,13 @@
 import type { GenerateConversationOptions } from "./conversation/types.js";
 import { type ResolvedModel, type Voice } from "./speech-provider.js";
-import type { ConversationResult, ConversationResultWithTimestamps } from "./speech-result.js";
-export { ConversationInputError, DialogueConstraintError, StitchUnsupportedError, } from "./conversation/errors.js";
+import type { ConversationResult, ConversationResultWithTimestamps, ConversationResultWithTurns } from "./speech-result.js";
+export { ConversationInputError, DialogueConstraintError, StitchUnsupportedError, TurnSplitError, type TurnSplitFailureReason, } from "./conversation/errors.js";
 export type { ConversationTurn, GenerateConversationOptions, } from "./conversation/types.js";
-export type { ConversationMetadata, ConversationResult, ConversationResultWithTimestamps, } from "./speech-result.js";
+export type { ConversationMetadata, ConversationPathKind, ConversationResult, ConversationResultWithTimestamps, ConversationResultWithTurns, ConversationStitchReason, ConversationTurnAudio, } from "./speech-result.js";
+export declare function generateConversation<V extends Voice = Voice, M extends string | ResolvedModel<V> | undefined = string | ResolvedModel<V> | undefined>(options: GenerateConversationOptions<V, M> & {
+    timestamps: true;
+    splitTurns: true;
+}): Promise<ConversationResultWithTurns>;
 export declare function generateConversation<V extends Voice = Voice, M extends string | ResolvedModel<V> | undefined = string | ResolvedModel<V> | undefined>(options: GenerateConversationOptions<V, M> & {
     timestamps: true;
 }): Promise<ConversationResultWithTimestamps>;

package/dist/generate-conversation.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"generate-conversation.d.ts","sourceRoot":"","sources":["../src/generate-conversation.ts"],"names":[],"mappings":"AAgBA,OAAO,KAAK,EAEV,2BAA2B,EAC5B,MAAM,yBAAyB,CAAC;AAyBjC,OAAO,EAGL,KAAK,aAAa,EAElB,KAAK,KAAK,EACX,MAAM,sBAAsB,CAAC;AAC9B,OAAO,KAAK,EAEV,kBAAkB,EAClB,gCAAgC,EACjC,MAAM,oBAAoB,CAAC;AAU5B,OAAO,EACL,sBAAsB,EACtB,uBAAuB,EACvB,sBAAsB,GACvB,MAAM,0BAA0B,CAAC;AAClC,YAAY,EACV,gBAAgB,EAChB,2BAA2B,GAC5B,MAAM,yBAAyB,CAAC;AACjC,YAAY,EACV,oBAAoB,EACpB,kBAAkB,EAClB,gCAAgC,GACjC,MAAM,oBAAoB,CAAC;AAkC5B,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS,EAEb,OAAO,EAAE,2BAA2B,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG;IAAE,UAAU,EAAE,IAAI,CAAA;CAAE,GAChE,OAAO,CAAC,gCAAgC,CAAC,CAAC;AAC7C,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS,EACb,OAAO,EAAE,2BAA2B,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAAC"}
+{"version":3,"file":"generate-conversation.d.ts","sourceRoot":"","sources":["../src/generate-conversation.ts"],"names":[],"mappings":"AAmBA,OAAO,KAAK,EAEV,2BAA2B,EAC5B,MAAM,yBAAyB,CAAC;AAwBjC,OAAO,EAGL,KAAK,aAAa,EAElB,KAAK,KAAK,EACX,MAAM,sBAAsB,CAAC;AAC9B,OAAO,KAAK,EAEV,kBAAkB,EAClB,gCAAgC,EAChC,2BAA2B,EAC5B,MAAM,oBAAoB,CAAC;AAU5B,OAAO,EACL,sBAAsB,EACtB,uBAAuB,EACvB,sBAAsB,EACtB,cAAc,EACd,KAAK,sBAAsB,GAC5B,MAAM,0BAA0B,CAAC;AAClC,YAAY,EACV,gBAAgB,EAChB,2BAA2B,GAC5B,MAAM,yBAAyB,CAAC;AACjC,YAAY,EACV,oBAAoB,EACpB,oBAAoB,EACpB,kBAAkB,EAClB,gCAAgC,EAChC,2BAA2B,EAC3B,wBAAwB,EACxB,qBAAqB,GACtB,MAAM,oBAAoB,CAAC;AAkC5B,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS,EAEb,OAAO,EAAE,2BAA2B,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG;IAC3C,UAAU,EAAE,IAAI,CAAC;IACjB,UAAU,EAAE,IAAI,CAAC;CAClB,GACA,OAAO,CAAC,2BAA2B,CAAC,CAAC;AACxC,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS,EAEb,OAAO,EAAE,2BAA2B,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG;IAAE,UAAU,EAAE,IAAI,CAAA;CAAE,GAChE,OAAO,CAAC,gCAAgC,CAAC,CAAC;AAC7C,wBAAgB,oBAAoB,CAClC,CAAC,SAAS,KAAK,GAAG,KAAK,EACvB,CAAC,SAAS,MAAM,GAAG,aAAa,CAAC,CAAC,CAAC,GAAG,SAAS,GAC3C,MAAM,GACN,aAAa,CAAC,CAAC,CAAC,GAChB,SAAS,EACb,OAAO,EAAE,2BAA2B,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAAC"}

package/dist/generate-conversation.js

@@ -1,9 +1,10 @@
 import pRetry from "p-retry";
-import { applySpeedToAudio, isSpeedActive, scaleTimestamps, validateSpeed, } from "./apply-speed.js";
+import { applySpeedToAudio, isSpeedActive, outputAfterSpeed, scaleTimestamps, validateSpeed, } from "./apply-speed.js";
 import { computeAudioDuration } from "./audio-duration.js";
 import { applyOptionalOutputConversion, sampleRateHintFrom, validateOutput, } from "./audio-output.js";
 import { mapWithConcurrency, resolveMaxConcurrency } from "./concurrency.js";
 import { chooseConversationPath } from "./conversation/dispatch.js";
+import { TurnSplitError } from "./conversation/errors.js";
 import { validateConversationInput } from "./conversation/validate.js";
 import { deriveTimestampsViaSTT } from "./derive-timestamps.js";
 import { NoSpeechGeneratedError, OutputConversionUnsupportedError, TimestampProviderRequiredError, TimestampValidationError, } from "./errors.js";
@@ -21,7 +22,7 @@ import { preprocessSpeechText } from "./text-preprocessing.js";
 import { deriveTimestampsViaProvider } from "./timestamp-alignment.js";
 import { finalizeTimestamps } from "./timestamp-finalization.js";
 // biome-ignore lint/performance/noBarrelFile: public entry point — re-export error classes so callers get fn + types + errors from one import
-export { ConversationInputError, DialogueConstraintError, StitchUnsupportedError, } from "./conversation/errors.js";
+export { ConversationInputError, DialogueConstraintError, StitchUnsupportedError, TurnSplitError, } from "./conversation/errors.js";
 const DEFAULT_GAP_MS = 300;
 const DEFAULT_MAX_RETRIES = 2;
 function requireValidTimestamps(args) {
@@ -84,12 +85,18 @@ export async function generateConversation(options) {
     for (const [index, resolved] of resolvedPerTurn.entries()) {
         validateInstructionSupport(resolved, combineInstructions(options.instructions, options.turns[index].instructions));
     }
-    const forceStitch = hasPerTurnSpeed ||
-        needsConversationStitchForMaxInputChars({
-            resolvedPerTurn,
-            turns: options.turns,
-            userMaxInputChars: options.maxInputChars,
-        });
+    const needsMaxInputCharsStitch = needsConversationStitchForMaxInputChars({
+        resolvedPerTurn,
+        turns: options.turns,
+        userMaxInputChars: options.maxInputChars,
+    });
+    let forceStitch;
+    if (hasPerTurnSpeed) {
+        forceStitch = "per-turn-speed";
+    }
+    else if (needsMaxInputCharsStitch) {
+        forceStitch = "max-input-chars";
+    }
     const path = chooseConversationPath({
         forceStitch,
         resolvedPerTurn,
@@ -97,14 +104,13 @@ export async function generateConversation(options) {
         output: options.output,
     });
     if (path.kind === "native") {
-        return await applySpeedToConversationResult({
-            result: await runNativeDispatch({
+        return await finishConversation({
+            run: await runNativeDispatch({
                 options,
                 resolved: path.resolved,
                 blocks: path.blocks,
             }),
-            speed: options.speed,
-            output: options.output,
+            options,
         });
     }
     // Lazy-load so native-only callers don't bundle pcm-concat / mediabunny.
@@ -119,7 +125,6 @@ export async function generateConversation(options) {
         maxConcurrency: resolveMaxConcurrency(options.maxConcurrency),
         maxInputChars: options.maxInputChars,
         maxRetries: options.maxRetries ?? DEFAULT_MAX_RETRIES,
-        output: options.output,
         volumeDbfs: options.volumeDbfs,
         abortSignal: options.abortSignal,
         headers: options.headers,
@@ -127,8 +132,6 @@ export async function generateConversation(options) {
         timestamps: options.timestamps ?? false,
         timestampProvider: options.timestampProvider,
         pronunciations: options.pronunciations,
-        // Defer output conversion to applySpeedToConversationResult to avoid encoding twice.
-        deferOutputConversion: isSpeedActive(options.speed),
     });
     if (stitched.audio.length === 0) {
         throw new NoSpeechGeneratedError(`${describeConversationModels(resolvedPerTurn)}: stitched conversation audio is empty.`, { reason: "provider_empty_response" });
@@ -140,11 +143,16 @@ export async function generateConversation(options) {
             audioDurationMs: stitched.metadata.audioDurationMs,
         }),
         perTurn: stitched.metadataPerTurn,
+        path: "stitch",
+        stitchReason: path.stitchReason,
     };
     let fallbackWarning;
     if (path.reason === "fallback-from-native") {
         fallbackWarning = `native dialogue unavailable because per-turn providerOptions are set; rendered via stitch (${options.turns.length} API calls instead of 1)`;
     }
+    else if (path.reason === "fallback-from-native-custom-voice") {
+        fallbackWarning = `native dialogue requires prebuilt voices; custom voices were rendered via stitch (${options.turns.length} generateSpeech calls) instead`;
+    }
     else if (path.reason === "fallback-from-native-oversized") {
         fallbackWarning = `native dialogue exceeds the provider's per-call limit and couldn't be split into voice-valid blocks; rendered via stitch (${options.turns.length} API calls instead of 1)`;
     }
@@ -157,20 +165,67 @@ export async function generateConversation(options) {
     const combinedWarnings = fallbackWarning
         ? [fallbackWarning, ...stitched.warnings]
         : stitched.warnings;
-    return await applySpeedToConversationResult({
-        result: {
-            audio: new DefaultGeneratedAudioFile({
-                data: stitched.audio,
-                mediaType: stitched.mediaType,
-            }),
-            metadata,
-            providerMetadata: { turns: stitched.providerMetadataPerTurn },
-            warnings: combinedWarnings.length > 0 ? [...combinedWarnings] : undefined,
-            timestamps: stitched.timestamps,
+    return await finishConversation({
+        run: {
+            result: {
+                audio: new DefaultGeneratedAudioFile({
+                    data: stitched.audio,
+                    mediaType: stitched.mediaType,
+                }),
+                metadata,
+                providerMetadata: { turns: stitched.providerMetadataPerTurn },
+                warnings: combinedWarnings.length > 0 ? [...combinedWarnings] : undefined,
+                timestamps: stitched.timestamps,
+            },
+            knownCutsSec: stitched.turnCutsSec,
         },
+        options,
+    });
+}
+// Runners return decodable audio; the requested output format is encoded exactly once, here.
+async function finishConversation(args) {
+    const { options, run } = args;
+    const speed = isSpeedActive(options.speed) ? options.speed : 1;
+    const output = speed === 1 ? options.output : outputAfterSpeed(options.output);
+    const result = await applySpeedToConversationResult({
+        result: run.result,
         speed: options.speed,
-        output: options.output,
     });
+    if (!options.splitTurns) {
+        return await encodeConversationAudio(result, output);
+    }
+    const { splitConversationTurns } = await import("./conversation/split-turns.js");
+    const split = await splitConversationTurns({
+        audio: result.audio.uint8Array,
+        knownCutsSec: run.knownCutsSec?.map((sec) => sec == null ? sec : sec / speed),
+        mediaType: result.audio.mediaType,
+        output,
+        timestamps: result.timestamps ?? [],
+        turnCount: options.turns.length,
+    });
+    const encoded = await encodeConversationAudio(result, output, {
+        audio: new Uint8Array(split.pcm.buffer, split.pcm.byteOffset, split.pcm.byteLength),
+        mediaType: `audio/pcm;rate=${split.sampleRate}`,
+    });
+    return { ...encoded, turns: split.turns };
+}
+async function encodeConversationAudio(result, output, 
+// Same audio already decoded, when the caller has it, to skip a second decode.
+source = {
+    audio: result.audio.uint8Array,
+    mediaType: result.audio.mediaType,
+}) {
+    if (!output) {
+        return result;
+    }
+    const converted = await applyOptionalOutputConversion({ ...source, output });
+    return {
+        ...result,
+        audio: new DefaultGeneratedAudioFile({
+            data: converted.audio,
+            mediaType: converted.mediaType,
+        }),
+    };
 }
 function needsConversationStitchForMaxInputChars(args) {
     let forceStitch = false;
@@ -220,6 +275,27 @@ function finalizeConversationTurnTimestamps(args) {
     });
     return projected;
 }
+// Force decodable PCM/WAV via getStitchOptions for normalization; if unavailable, emit the provider's mixed audio and warn.
+function resolveNativeStitchOptions(options, resolved) {
+    const stitchOpts = resolved.provider.getStitchOptions?.(resolved.modelId, {
+        sampleRate: sampleRateHintFrom(options.output),
+    });
+    if (stitchOpts) {
+        return { stitchOpts, warnings: [] };
+    }
+    if (options.splitTurns) {
+        throw new TurnSplitError({ reason: "undecodable_audio" });
+    }
+    if (options.output) {
+        throw new OutputConversionUnsupportedError(`${resolved.provider.id}/${resolved.modelId}`);
+    }
+    return {
+        stitchOpts,
+        warnings: [
+            `${resolved.provider.id}/${resolved.modelId}: native dialogue path returns the provider's mixed audio without volume normalization (no decodable PCM/WAV mode).`,
+        ],
+    };
+}
 async function runNative(args) {
     const { options, resolved, maxRetries } = args;
     const start = performance.now();
@@ -227,17 +303,7 @@ async function runNative(args) {
         throw new Error(`generateConversation: ${resolved.provider.id}/${resolved.modelId} dispatched to native but generateDialogue missing`);
     }
     const generateDialogue = resolved.provider.generateDialogue.bind(resolved.provider);
-    // Force decodable PCM/WAV via getStitchOptions for normalization; if unavailable, emit the provider's mixed audio and warn.
-    const stitchOpts = resolved.provider.getStitchOptions?.(resolved.modelId, {
-        sampleRate: sampleRateHintFrom(options.output),
-    });
-    const warnings = [];
-    if (!stitchOpts) {
-        warnings.push(`${resolved.provider.id}/${resolved.modelId}: native dialogue path returns the provider's mixed audio without volume normalization (no decodable PCM/WAV mode).`);
-    }
-    if (options.output && !stitchOpts) {
-        throw new OutputConversionUnsupportedError(`${resolved.provider.id}/${resolved.modelId}`);
-    }
+    const { stitchOpts, warnings } = resolveNativeStitchOptions(options, resolved);
     // Stitch options must win — caller-supplied response_format would break the decoder.
     const dialogueProviderOptions = stitchOpts
         ? { ...options.providerOptions, ...stitchOpts.providerOptions }
@@ -289,7 +355,7 @@ async function runNative(args) {
     });
     const computedDuration = await computeAudioDuration(audio.uint8Array, outputMediaType);
     const audioDurationMs = computedDuration ?? result.audioDurationMs;
-    const { timestamps: rawTimestamps, warnings: attributionWarnings } = await resolveNativeDialogueTimestamps({
+    const rawTimestamps = await resolveNativeDialogueTimestamps({
         requestTimestamps,
         nativeTimestamps: result.timestamps,
         hasNativeTimestamps: hasNativeDialogueTimestamps,
@@ -311,42 +377,30 @@ async function runNative(args) {
         timestamps: projection.timestamps,
         turnTexts: substitutedTurns.map((turn) => turn.originalText),
     });
-    // Defer output conversion to applySpeedToConversationResult when top-level speed
-    // is active — otherwise we'd encode here and re-encode in the stretch step.
-    const deferOutput = isSpeedActive(options.speed);
-    const converted = await applyOptionalOutputConversion({
-        audio: audio.uint8Array,
-        mediaType: outputMediaType,
-        output: deferOutput ? undefined : options.output,
-    });
-    const finalAudio = options.output && !deferOutput
-        ? new DefaultGeneratedAudioFile({
-            data: converted.audio,
-            mediaType: converted.mediaType,
-        })
-        : audio;
     const inputChars = options.turns.reduce((n, t) => n + t.text.length, 0);
     const metadata = {
         latencyMs,
         inputChars,
         ...(audioDurationMs != null && { audioDurationMs }),
+        path: "native",
     };
     const preprocessingWarnings = substitutedTurns.flatMap((turn) => turn.warnings);
     const mergedWarningList = [
         ...warnings,
         ...preprocessingWarnings,
-        ...attributionWarnings,
         ...(projection.estimatedBoundaries
             ? [PRONUNCIATION_TIMESTAMP_ESTIMATE_WARNING]
             : []),
     ];
     const mergedWarnings = mergedWarningList.length > 0 ? mergedWarningList : undefined;
     return {
-        audio: finalAudio,
-        metadata,
-        providerMetadata: result.providerMetadata,
-        warnings: mergedWarnings,
-        timestamps,
+        result: {
+            audio,
+            metadata,
+            providerMetadata: result.providerMetadata,
+            warnings: mergedWarnings,
+            timestamps,
+        },
     };
 }
 async function runNativeDispatch(args) {
@@ -429,7 +483,7 @@ async function runNativeSplit(args) {
             mediaType: stitchOpts.mediaType,
         }).uint8Array;
         const segment = await decodeAudioToPcm16(blockAudio, stitchOpts.mediaType);
-        const { timestamps, warnings } = await resolveNativeDialogueTimestamps({
+        const timestamps = await resolveNativeDialogueTimestamps({
             requestTimestamps,
             nativeTimestamps: result.timestamps,
             hasNativeTimestamps,
@@ -441,20 +495,29 @@ async function runNativeSplit(args) {
             audioDurationMs: (segment.pcm.length / segment.sampleRate) * 1000,
             substitutedTurnTexts: blockTurns.map((t) => t.canonicalText),
             timestampProvider: options.timestampProvider,
-            pcmSegment: segment,
         });
         return {
             segment,
             timestamps,
-            warnings,
             providerMetadata: result.providerMetadata,
         };
     }, { signal: options.abortSignal });
-    const { concatPcmToWav, dbfsToInt16Rms, normalizeRms, stitchTargetRate } = await import("./conversation/pcm-concat.js");
+    const { concatPcmToWavWithRanges, dbfsToInt16Rms, gapMidpointsSec, normalizeRms, stitchTargetRate, } = await import("./conversation/pcm-concat.js");
     const segments = perBlock.map((p) => p.segment);
     const leveled = normalizeRms(segments, options.volumeDbfs == null ? undefined : dbfsToInt16Rms(options.volumeDbfs));
     const targetSampleRate = stitchTargetRate(leveled);
-    const wav = await concatPcmToWav(leveled, { gapMs, targetSampleRate });
+    const { wav, ranges } = await concatPcmToWavWithRanges(leveled, {
+        gapMs,
+        targetSampleRate,
+    });
+    // Block joins are the only boundaries the SDK cut itself; boundaries inside a block stay undefined.
+    const knownCutsSec = [];
+    for (const [b, cutSec] of gapMidpointsSec(ranges, targetSampleRate).entries()) {
+        const lastTurn = blocks[b].at(-1);
+        if (lastTurn != null) {
+            knownCutsSec[lastTurn] = cutSec;
+        }
+    }
     const timestampProjection = requestTimestamps
         ? composeBlockTimestamps({
             perBlock,
@@ -465,15 +528,9 @@ async function runNativeSplit(args) {
         })
         : { estimatedBoundaries: false, timestamps: undefined };
     const timestamps = timestampProjection.timestamps;
-    const deferOutput = isSpeedActive(options.speed);
-    const converted = await applyOptionalOutputConversion({
-        audio: wav,
-        mediaType: "audio/wav",
-        output: deferOutput ? undefined : options.output,
-    });
     const finalAudio = new DefaultGeneratedAudioFile({
-        data: converted.audio,
-        mediaType: converted.mediaType,
+        data: wav,
+        mediaType: "audio/wav",
     });
     // Derive duration from the PCM sample counts (resampled to the stitch rate plus gaps)
     // instead of re-decoding the merged audio — matches the stitch path.
@@ -486,20 +543,23 @@ async function runNativeSplit(args) {
         latencyMs: Math.round(performance.now() - start),
         inputChars,
         ...(audioDurationMs != null && { audioDurationMs }),
+        path: "native-split",
     };
     const warnings = [
         ...substitutedTurns.flatMap((turn) => turn.warnings),
-        ...perBlock.flatMap((p) => p.warnings),
         ...(timestampProjection.estimatedBoundaries
             ? [PRONUNCIATION_TIMESTAMP_ESTIMATE_WARNING]
             : []),
     ];
     return {
-        audio: finalAudio,
-        metadata,
-        providerMetadata: { blocks: perBlock.map((p) => p.providerMetadata) },
-        warnings: warnings.length > 0 ? [...warnings] : undefined,
-        timestamps,
+        result: {
+            audio: finalAudio,
+            metadata,
+            providerMetadata: { blocks: perBlock.map((p) => p.providerMetadata) },
+            warnings: warnings.length > 0 ? [...warnings] : undefined,
+            timestamps,
+        },
+        knownCutsSec,
     };
 }
 function composeBlockTimestamps(args) {
@@ -538,7 +598,7 @@ function composeBlockTimestamps(args) {
 }
 async function resolveNativeDialogueTimestamps(args) {
     if (!args.requestTimestamps) {
-        return { timestamps: undefined, warnings: [] };
+        return;
     }
     // Either use native flat timestamps, or derive via STT fallback.
     let flatTimestamps;
@@ -582,36 +642,24 @@ async function resolveNativeDialogueTimestamps(args) {
         text: args.substitutedTurnTexts.join(" "),
         timestamps: flatTimestamps,
     });
-    const { detectSilenceGaps } = await import("./conversation/silence-detection.js");
     const { attributeTimestamps } = await import("./conversation/attribute-timestamps.js");
-    let silenceGaps = [];
-    try {
-        const segment = args.pcmSegment ??
-            (await (await import("./audio-decode.js")).decodeAudioToPcm16(args.audio, args.mediaType));
-        const gaps = detectSilenceGaps(segment.pcm, {
-            sampleRate: segment.sampleRate,
-            minDurationMs: 150,
-        });
-        silenceGaps = gaps;
-    }
-    catch {
-        // Decoder couldn't read the audio (e.g., compressed format we can't decode locally).
-        // Tier 1 will be skipped; dispatcher falls through to Tier 2/3.
-    }
-    const result = attributeTimestamps({
+    // Validated words cover the joined turn text exactly, so an exact partition always exists.
+    const attributed = attributeTimestamps({
         timestamps: flatTimestamps,
         turnTexts: args.substitutedTurnTexts,
-        silenceGaps,
     });
-    return {
-        timestamps: finalizeConversationTurnTimestamps({
-            audioDurationMs: args.audioDurationMs,
+    if (!attributed) {
+        throw new TimestampValidationError({
+            reason: "transcript_mismatch",
             source: args.ttsModel,
-            timestamps: result.timestamps,
-            turnTexts: args.substitutedTurnTexts,
-        }),
-        warnings: result.warnings,
-    };
+        });
+    }
+    return finalizeConversationTurnTimestamps({
+        audioDurationMs: args.audioDurationMs,
+        source: args.ttsModel,
+        timestamps: attributed,
+        turnTexts: args.substitutedTurnTexts,
+    });
 }
 async function buildNativeAudio(args) {
     let audioBytes = args.result.audio;
@@ -663,7 +711,7 @@ function buildSubstitutedTurns(turns, resolved, ruleMap) {
     });
 }
 async function applySpeedToConversationResult(args) {
-    const { result, speed, output } = args;
+    const { result, speed } = args;
     if (!isSpeedActive(speed)) {
         return result;
     }
@@ -671,7 +719,8 @@ async function applySpeedToConversationResult(args) {
         audio: result.audio.uint8Array,
         mediaType: result.audio.mediaType,
         speed,
-        output,
+        // finishConversation encodes the requested format once, after any split.
+        output: { format: "wav" },
     });
     const newAudio = new DefaultGeneratedAudioFile({
         data: stretched.audio,