arcane-os 0.5.3 → 0.5.5

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## 0.5.5
+
+- Added configurable complete text-to-speech stream segmentation through
+  `AI.configureTTSSegmentation()`. Applications can speak on Unicode punctuation
+  or a whole-word cadence without discarding any text, while apostrophes,
+  commas, and hyphens that join Unicode letters or numbers remain intact.
+  Synthesis and playback stay sequential, and mute or cancellation still stops
+  the complete queue.
+
+## 0.5.4
+
+- Reduced Browser-WASM HTTP Range parts from roughly 128 MB to roughly 4 MB
+  and raised the deterministic part ceiling, so a refresh re-fetches only the
+  small in-flight parts while every completed part remains reusable. An
+  incomplete 0.5.3 cache keeps its completed legacy parts and subdivides only
+  the missing legacy intervals, preserving existing progress during adoption.
+- Kept all built-in audio processing on device: Whisper owns transcription and
+  Kokoro owns speech synthesis. Legacy OpenAI audio selections migrate to those
+  local routes, non-local speech configurations are rejected, and the retired
+  OpenAI audio credential and endpoints are no longer used.
+
 ## 0.5.3
 
 - Preserved every independently completed Browser-WASM HTTP Range part and

package/README.md

@@ -19,7 +19,7 @@ version-locked SDK runtime, while an integrated Arcane checkout uses its live
 `arcane/` runtime. Both profiles preserve the same app URLs, theme, packaging,
 event, cancellation, and browser run contracts.
 
-This checkout defines the `0.5.3` SDK contract. Applications pin one exact npm
+This checkout defines the `0.5.5` SDK contract. Applications pin one exact npm
 version and lockfile; registry state is deliberately not baked into application
 artifacts.
 
@@ -88,10 +88,10 @@ OpenAI-compatible chat-completion requests to
 `globalThis.arcane.config.twinCloud.accessKey`. The established `ai.license`
 property and internal `OPENAI` route identifier remain compatibility aliases;
 applications should present the service and credential as **TWiN Cloud** and
-**TWiN access key**. The TWiN key is used only for remote LLM chat. Existing
-OpenAI speech routes remain separate and read only their legacy
-`globalThis.arcane.config.openAI.apiKey`; the SDK never sends a TWiN key to
-those speech endpoints.
+**TWiN access key**. The TWiN key is used only for remote LLM chat. Audio stays
+on device: Whisper (`LOCAL_SPEACH` / `whisper-small`) owns transcription and
+Kokoro (`LOCAL_SPEACH` / `kokoro`) owns speech synthesis. Neither audio route
+uses the TWiN key, and no OpenAI audio key is required.
 
 ## Browser-local AI
 
@@ -124,11 +124,14 @@ Optional hardening is inactive unless the caller explicitly selects
 `secure:true`. The DBOPFS store downloads ordered members with one bounded
 parallel transfer pool and returns them in descriptor order. Completed shards
 survive an interrupted attempt and only missing shard or Range-part work is
-fetched on retry. A
-monolithic file uses up to `downloadConcurrency` active workers (four by
-default) over up to 16 deterministic, resumable HTTP range parts when the server
-confirms `206` responses and the total comes from `Content-Range` or, when that
-header is not exposed, optional declared `bytes`.
+fetched on retry. A monolithic file uses up to `downloadConcurrency` active
+workers (four by default) over deterministic, resumable HTTP range parts of
+roughly 4 MB each, up to 4,096 parts, when the server confirms `206` responses
+and the total comes from `Content-Range` or, when that header is not exposed,
+optional declared `bytes`.
+An incomplete cache written by 0.5.3 keeps its completed coarse parts and
+subdivides only the missing intervals into the smaller current parts, so an SDK
+update preserves existing progress while reducing later refresh loss.
 When a followed redirect turns the first Range probe into `200`, the SDK probes
 the final URL directly before reusing that original response as the single-fetch
 fallback. Completed range parts survive restart and are presented to Wllama as
@@ -150,6 +153,24 @@ must supply each runtime/model selection explicitly. Speech roles
 load, cancel, unload, fail, and recover independently, so speech failure never
 silently falls back or prevents text chat.
 
+Applications that need faster spoken-response onset can configure the shared
+TTS stream without taking over synthesis or playback:
+
+```js
+ai.configureTTSSegmentation({
+    punctuation:'any',
+    wordCadence:4
+});
+```
+
+The compatibility default remains sentence punctuation with no word cadence.
+The configured stream preserves every character and punctuation mark, chooses
+the earliest complete boundary, and keeps synthesis and playback sequential.
+Any-punctuation mode recognizes boundaries without requiring whitespace while
+keeping apostrophes, commas, and hyphens that join Unicode letters or numbers
+inside the same segment.
+Mute, stop, provider transition, and cancellation still govern the whole queue.
+
 The SDK runtime also owns `DBOPFSDocumentLibrary`,
 `DocumentLexicalSearch`, and `PersistentAIChatSession`. Document bootstrap is
 schema-driven and explicit; chat never searches a corpus unless the app wires
@@ -165,7 +186,7 @@ uses the same controller for automatic memory extraction.
 Create a new repository-shaped Arcane application with the exact stable SDK:
 
 ```bash
-npx arcane-os@0.5.3 new my-app --path ./my-app --target portable --git
+npx arcane-os@0.5.5 new my-app --path ./my-app --target portable --git
 cd my-app
 npm install
 npm run check
@@ -176,7 +197,7 @@ To enroll an existing repository, install the exact SDK and initialize only
 missing Arcane files:
 
 ```bash
-npm install --save-dev --save-exact arcane-os@0.5.3
+npm install --save-dev --save-exact arcane-os@0.5.5
 npm exec -- arcane init my-app --target portable
 ```
 
@@ -192,7 +213,7 @@ npm exec -- arcane-os targets
 No global SDK install or standalone Arcane CLI is required. The application
 repository's exact npm dependency and lockfile own the CLI and toolchain version.
 
-Use `npx arcane-os@0.5.3` for the initial bootstrap because it names this npm
+Use `npx arcane-os@0.5.5` for the initial bootstrap because it names this npm
 package explicitly; bare `npx arcane` outside an installed project could resolve
 a different package. Both installed commands invoke the same headless toolchain.
 Project-local npm scripts use the SDK pinned by that app's `package-lock.json`,
@@ -213,7 +234,7 @@ node ./bin/arcane.mjs new local-app --path ../local-app --target portable --git
 
 # From the generated app repository
 cd ../local-app
-npm install --save-dev --save-exact ../arcane-os-sdk/arcane-os-0.5.3.tgz
+npm install --save-dev --save-exact ../arcane-os-sdk/arcane-os-0.5.5.tgz
 npm run check
 npm ci
 ```
@@ -223,7 +244,7 @@ same location. The lockfile retains the selected package dependency while
 Arcane uses the installed package name and version. Local directory `file:` dependencies are not
 accepted because npm may install them as links; use a packed `.tgz`. A GitHub
 runner also needs that tarball at the locked path. After publication, replace
-the local declaration with the exact `arcane-os@0.5.3` registry package and
+the local declaration with the exact `arcane-os@0.5.5` registry package and
 commit the regenerated lock.
 
 Generated repositories use `npm ci --ignore-scripts` in CI. Run dependency
@@ -361,7 +382,7 @@ package installation, or assertions.
 
 ## Current target support
 
-Version `0.5.3` exposes one browser target and five explicitly paired
+Version `0.5.5` exposes one browser target and five explicitly paired
 native development targets: a non-runnable portable directory, a
 Windows x64 unsigned-local-test EXE bundle, Linux x64 and Linux ARM64
 unsigned-local-test DEBs, and an Android development-signed APK. The

package/browser-runtime/ai/browser-wasm-llm-provider.mjs

@@ -29,8 +29,10 @@ const MODEL_LOAD_HEARTBEAT_MS = 5_000;
 const DEFAULT_MODEL_DOWNLOAD_CONCURRENCY = 4;
 const MODEL_DOWNLOAD_PROGRESS_INTERVAL_MS = 250;
 const MODEL_DOWNLOAD_SPEED_WINDOW_MS = 5_000;
-const MODEL_DOWNLOAD_MAX_RANGE_PARTS = 16;
-const MODEL_DOWNLOAD_TARGET_RANGE_BYTES = 128_000_000;
+const MODEL_DOWNLOAD_MAX_RANGE_PARTS = 4_096;
+const MODEL_DOWNLOAD_TARGET_RANGE_BYTES = 4_000_000;
+const LEGACY_MODEL_DOWNLOAD_MAX_RANGE_PARTS = 16;
+const LEGACY_MODEL_DOWNLOAD_TARGET_RANGE_BYTES = 128_000_000;
 const INTEL_VENDOR_ID = 0x8086;
 const CAPABILITY_POLICY_PROTOCOL = "arcane-ai-browser-capability-policy/1";
 let highPerformanceGpuNoticeShown = false;
@@ -103,10 +105,10 @@ function downloadConcurrencyValue(value) {
   return value;
 }
 
-function modelHttpRanges(total) {
+function modelHttpRangesFor(total, maximumParts, targetPartBytes) {
   const count = Math.min(
-    MODEL_DOWNLOAD_MAX_RANGE_PARTS,
-    Math.ceil(total / MODEL_DOWNLOAD_TARGET_RANGE_BYTES),
+    maximumParts,
+    Math.ceil(total / targetPartBytes),
     total,
   );
   const width = Math.floor(total / count);
@@ -122,6 +124,47 @@ function modelHttpRanges(total) {
   return completeValue(ranges);
 }
 
+function modelHttpRanges(total) {
+  return modelHttpRangesFor(
+    total,
+    MODEL_DOWNLOAD_MAX_RANGE_PARTS,
+    MODEL_DOWNLOAD_TARGET_RANGE_BYTES,
+  );
+}
+
+function modelHttpSubranges(range) {
+  const length = range.end - range.start + 1;
+  const maximumParts = Math.floor(
+    MODEL_DOWNLOAD_MAX_RANGE_PARTS / LEGACY_MODEL_DOWNLOAD_MAX_RANGE_PARTS,
+  );
+  return completeValue(modelHttpRangesFor(
+    length,
+    maximumParts,
+    MODEL_DOWNLOAD_TARGET_RANGE_BYTES,
+  ).map((part) => completeValue({
+    start: range.start + part.start,
+    end: range.start + part.end,
+    total: range.total,
+  })));
+}
+
+function modelHttpRangePlans(total) {
+  const current = modelHttpRanges(total);
+  const legacy = modelHttpRangesFor(
+    total,
+    LEGACY_MODEL_DOWNLOAD_MAX_RANGE_PARTS,
+    LEGACY_MODEL_DOWNLOAD_TARGET_RANGE_BYTES,
+  );
+  if (
+    current.length === legacy.length
+    && current.every((range, index) => (
+      range.start === legacy[index].start
+      && range.end === legacy[index].end
+    ))
+  ) return completeValue([current]);
+  return completeValue([current, legacy]);
+}
+
 function progressClock() {
   return typeof globalThis.performance?.now === "function"
     ? globalThis.performance.now()
@@ -930,7 +973,7 @@ export function createDbopfsModelStore({
     } = {},
   ) {
     const transport = BROWSER_MODEL_SOURCE_TRANSPORTS.get(source);
-    const ranges = modelHttpRanges(total);
+    const ranges = await rangePlanForInstall(name, total, signal);
     const partFiles = new Array(ranges.length);
     const pendingRangeIndexes = [];
     for (let rangeIndex = 0; rangeIndex < ranges.length; rangeIndex += 1) {
@@ -1003,16 +1046,31 @@ export function createDbopfsModelStore({
           writable = null;
           partFiles[rangeIndex] = await storedModelFile(partName);
         } catch (error) {
-          retainFailure(error);
-          await cancelOpenedDownload(opened, error);
+          let transferError = error;
+          if (downloadSignal.aborted
+            && error?.code !== "ARCANE_AI_REQUEST_ABORTED") {
+            try {
+              throwIfAborted(downloadSignal, "install");
+            } catch (abortError) {
+              transferError = abortError;
+            }
+          } else if (!error?.code) {
+            transferError = fail(
+              "ARCANE_AI_MODEL_DOWNLOAD_FAILED",
+              "The model download failed while reading an HTTP byte range.",
+              error,
+            );
+          }
+          retainFailure(transferError);
+          await cancelOpenedDownload(opened, transferError);
           try {
-            await writable?.abort?.(error);
+            await writable?.abort?.(transferError);
           } catch {
             // The range part is removed below even if its writable already closed.
           }
           await removeEntry(partName).catch(() => undefined);
           progress?.discardBytes(received);
-          throw error;
+          throw transferError;
         } finally {
           progress?.endTransfer({ publishProgress: failure === null });
         }
@@ -1118,16 +1176,93 @@ export function createDbopfsModelStore({
     return names;
   }
 
+  async function rangePlanForInstall(
+    modelName,
+    total,
+    signal,
+    knownNames = undefined,
+  ) {
+    const plans = modelHttpRangePlans(total);
+    const existingNames = knownNames === undefined
+      ? await memberRangePartNames(modelName)
+      : knownNames;
+    const available = existingNames === null
+      ? null
+      : existingNames instanceof Set
+        ? existingNames
+        : new Set(existingNames);
+    if (available?.size === 0) return plans[0];
+    if (plans.length === 1) return plans[0];
+    const completedParts = new Map();
+
+    async function completedPart(range) {
+      throwIfAborted(signal, "install");
+      const partName = rangePartName(modelName, range);
+      if (available && !available.has(partName)) return null;
+      if (completedParts.has(partName)) return completedParts.get(partName);
+      const partFile = await file(partName);
+      const expected = range.end - range.start + 1;
+      const completed = partFile?.size === expected ? partFile : null;
+      completedParts.set(partName, completed);
+      return completed;
+    }
+
+    const completedLegacyRanges = new Set();
+    for (const range of plans[1]) {
+      if (await completedPart(range)) completedLegacyRanges.add(range);
+    }
+    const hybrid = completeValue(plans[1].flatMap((range) => (
+      completedLegacyRanges.has(range)
+        ? [range]
+        : modelHttpSubranges(range)
+    )));
+    let selected = plans[0];
+    let selectedBytes = 0;
+    let selectedLastModified = 0;
+    for (const ranges of [plans[0], hybrid]) {
+      let completedBytes = 0;
+      let lastModified = 0;
+      for (const range of ranges) {
+        const partFile = await completedPart(range);
+        if (!partFile) continue;
+        completedBytes += range.end - range.start + 1;
+        if (Number.isFinite(partFile.lastModified)) {
+          lastModified = Math.max(lastModified, partFile.lastModified);
+        }
+      }
+      if (
+        completedBytes > selectedBytes
+        || (
+          completedBytes === selectedBytes
+          && completedBytes > 0
+          && lastModified > selectedLastModified
+        )
+      ) {
+        selected = ranges;
+        selectedBytes = completedBytes;
+        selectedLastModified = lastModified;
+      }
+    }
+    return selected;
+  }
+
   async function removeMemberRangeParts(modelName, {
     except = null,
     total = null,
   } = {}) {
     const existingNames = await memberRangePartNames(modelName);
-    const names = existingNames ?? (
-      Number.isSafeInteger(total) && total > 0
-        ? modelHttpRanges(total).map((range) => rangePartName(modelName, range))
-        : []
-    );
+    let names = existingNames;
+    if (names === null && Number.isSafeInteger(total) && total > 0) {
+      const plans = modelHttpRangePlans(total);
+      const hybridRanges = plans.length === 1
+        ? []
+        : plans[1].flatMap((range) => modelHttpSubranges(range));
+      names = [...new Set([
+        ...plans.flat(),
+        ...hybridRanges,
+      ].map((range) => rangePartName(modelName, range)))];
+    }
+    names ??= [];
     const removed = [];
     for (const name of names) {
       if (except?.has(name)) continue;
@@ -1211,9 +1346,10 @@ export function createDbopfsModelStore({
     for (let memberIndex = 0; memberIndex < members.length; memberIndex += 1) {
       const total = members[memberIndex].bytes;
       if (!Number.isSafeInteger(total) || total <= 0) continue;
-      for (const range of modelHttpRanges(total)) {
-        removed.push(await removeEntry(rangePartName(names.models[memberIndex].name, range)));
-      }
+      removed.push(await removeMemberRangeParts(
+        names.models[memberIndex].name,
+        { total },
+      ));
     }
     return removed.some(Boolean);
   }
@@ -1364,12 +1500,18 @@ export function createDbopfsModelStore({
     return modelFiles.every(Boolean) ? modelFiles : null;
   }
 
-  async function storedRangeCandidate(modelName, total, signal) {
+  async function storedRangeCandidateForPlan(
+    modelName,
+    ranges,
+    signal,
+    availableNames = null,
+  ) {
     const partFiles = [];
     let lastModified = 0;
-    for (const range of modelHttpRanges(total)) {
+    for (const range of ranges) {
       throwIfAborted(signal, "install");
       const partName = rangePartName(modelName, range);
+      if (availableNames && !availableNames.has(partName)) return null;
       const partFile = await file(partName);
       if (!partFile) return null;
       const expected = range.end - range.start + 1;
@@ -1385,26 +1527,53 @@ export function createDbopfsModelStore({
     return completeValue({
       file: new Blob(partFiles, { type: "application/octet-stream" }),
       lastModified,
-      total,
+      ranges,
+      total: ranges[0].total,
     });
   }
 
+  async function storedRangeCandidate(
+    modelName,
+    total,
+    signal,
+    availableNames = null,
+  ) {
+    const ranges = await rangePlanForInstall(
+      modelName,
+      total,
+      signal,
+      availableNames,
+    );
+    return storedRangeCandidateForPlan(
+      modelName,
+      ranges,
+      signal,
+      availableNames,
+    );
+  }
+
   async function storedRangeMember(member, modelName, signal) {
+    const names = await memberRangePartNames(modelName);
+    const availableNames = names === null ? null : new Set(names);
     const declaredTotal = Number.isSafeInteger(member.bytes) && member.bytes > 0
       ? member.bytes
       : null;
     if (declaredTotal !== null) {
-      const declared = await storedRangeCandidate(modelName, declaredTotal, signal);
+      const declared = await storedRangeCandidate(
+        modelName,
+        declaredTotal,
+        signal,
+        availableNames,
+      );
       if (declared) {
         await removeStaleRangePartsAfterCompletion(
           modelName,
-          modelHttpRanges(declaredTotal),
+          declared.ranges,
         );
         return declared.file;
       }
     }
 
-    const names = await memberRangePartNames(modelName);
     if (names === null) return null;
     const discoveredTotals = new Set();
     for (const name of names) {
@@ -1414,7 +1583,12 @@ export function createDbopfsModelStore({
     let selected = null;
     const totals = [...discoveredTotals].sort((left, right) => right - left);
     for (const total of totals) {
-      const candidate = await storedRangeCandidate(modelName, total, signal);
+      const candidate = await storedRangeCandidate(
+        modelName,
+        total,
+        signal,
+        availableNames,
+      );
       if (
         candidate
         && (
@@ -1426,7 +1600,7 @@ export function createDbopfsModelStore({
     if (selected === null) return null;
     await removeStaleRangePartsAfterCompletion(
       modelName,
-      modelHttpRanges(selected.total),
+      selected.ranges,
     );
     return selected.file;
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "arcane-os",
-  "version": "0.5.3",
+  "version": "0.5.5",
   "description": "Arcane OS JavaScript SDK, project-local CLI, browser runtime, and repository-portable application packager.",
   "type": "module",
   "main": "./src/index.mjs",

package/runtime/arcane/modules/AI.js

@@ -17,6 +17,16 @@ const completeValue=(value)=>value;
 
 let credentials='include';
 const LEGACY_TTS_RESPONSE_FORMAT='opus';
+const DEFAULT_TTS_SEGMENTATION={
+    punctuation:'sentence',
+    wordCadence:null
+};
+const TTS_PUNCTUATION_MODES=new Set(['sentence','any','none']);
+// A complete punctuation run is a boundary unless the whole run consists of
+// apostrophe, comma, or dash punctuation joining Unicode letters or numbers.
+// The streaming expression leaves a trailing joining run for the next chunk.
+const TTS_ANY_PUNCTUATION=/(?<!\p{P})(?:(?<![\p{L}\p{N}])\p{P}+(?!\p{P})(?=[\s\S])|\p{P}+(?!\p{P})(?=[^\p{L}\p{N}])|\p{P}*[^\P{P}'\u2019\uFF07,\u060C\u3001\uFF0C\p{Pd}]\p{P}*)(?!\p{P})/gu;
+const TTS_ANY_PUNCTUATION_AT_END=/(?<!\p{P})(?:(?<![\p{L}\p{N}])\p{P}+|\p{P}+(?![\p{L}\p{N}])|\p{P}*[^\P{P}'\u2019\uFF07,\u060C\u3001\uFF0C\p{Pd}]\p{P}*)(?!\p{P})/gu;
 credentials='omit';
 
 const LEGACY_AI_SERVICES=new Set(['OPENAI','OLLAMA','LOCAL_SPEACH']);
@@ -142,6 +152,38 @@ function isPlainAIRecord(value){
     return prototype===Object.prototype||prototype===null;
 }
 
+function normalizeTTSSegmentation(value,current=DEFAULT_TTS_SEGMENTATION){
+    if(value===undefined||value===null){
+        return {...current};
+    }
+    if(!isPlainAIRecord(value)){
+        throw new TypeError('TTS segmentation must be a plain object.');
+    }
+
+    const punctuation=value.punctuation===undefined
+        ?current.punctuation
+        :value.punctuation;
+    if(!TTS_PUNCTUATION_MODES.has(punctuation)){
+        throw new TypeError(
+            'TTS segmentation punctuation must be sentence, any, or none.'
+        );
+    }
+
+    const wordCadence=value.wordCadence===undefined
+        ?current.wordCadence
+        :value.wordCadence;
+    if(
+        wordCadence!==null
+        &&(!Number.isSafeInteger(wordCadence)||wordCadence<1)
+    ){
+        throw new RangeError(
+            'TTS segmentation wordCadence must be null or a positive integer.'
+        );
+    }
+
+    return {punctuation,wordCadence};
+}
+
 function requireAIToolMessageSchemas(value,label='AI tools'){
     if(!Array.isArray(value)){
         throw aiStructuralError(
@@ -922,12 +964,10 @@ class AI {
             OPENAI: 'https://inference.do-ai.run/v1'
         },
         sttURL: {
-            LOCAL_SPEACH: 'http://127.0.0.1:8011/v1',
-            OPENAI:       'https://api.openai.com/v1'
+            LOCAL_SPEACH: 'http://127.0.0.1:8011/v1'
         },
         ttsURL: {
-            LOCAL_SPEACH: 'http://127.0.0.1:8011/v1',
-            OPENAI:       'https://api.openai.com/v1'
+            LOCAL_SPEACH: 'http://127.0.0.1:8011/v1'
         },
     }
 
@@ -936,12 +976,10 @@ class AI {
             OPENAI: '/chat/completions'
         },
         stt: {
-            LOCAL_SPEACH: '/audio/transcriptions',
-            OPENAI:       '/audio/transcriptions'
+            LOCAL_SPEACH: '/audio/transcriptions'
         },
         tts: {
-            LOCAL_SPEACH: '/audio/speech',
-            OPENAI:       '/audio/speech'
+            LOCAL_SPEACH: '/audio/speech'
         }
     }
 
@@ -950,12 +988,10 @@ class AI {
     }
 
     #sttModels = {
-        OPENAI:       'whisper-1',
         LOCAL_SPEACH: 'whisper-small'
     }
 
     #ttsModels = {
-        OPENAI:       'gpt-4o-mini-tts',
         LOCAL_SPEACH: 'kokoro'
     }
 
@@ -974,17 +1010,8 @@ class AI {
         };
     }
 
-    get #legacyOpenAISpeechKey(){
-        const value=globalThis.arcane?.config?.openAI?.apiKey;
-        return typeof value==='string'?value:'';
-    }
-
     get #ttsHeaders(){
         return {
-            OPENAI: {
-                'Content-Type': 'application/json',
-                'Authorization': `Bearer ${this.#legacyOpenAISpeechKey}`
-            },
             LOCAL_SPEACH: {
                 'Content-Type': 'application/json',
             }
@@ -993,9 +1020,6 @@ class AI {
 
     get #sttHeaders(){
         return {
-            OPENAI: {
-                'Authorization': `Bearer ${this.#legacyOpenAISpeechKey}`,
-            },
             LOCAL_SPEACH: {}
         };
     }
@@ -1048,11 +1072,11 @@ class AI {
 
         const preferences=[
             llmService||'OPENAI',
-            sttService||'OPENAI',
-            ttsService||'OPENAI',
+            sttService||'LOCAL_SPEACH',
+            ttsService||'LOCAL_SPEACH',
             model||'OPENAI',
-            modelTTS||'OPENAI',
-            modelSTT||'OPENAI'
+            modelTTS||'LOCAL_SPEACH',
+            modelSTT||'LOCAL_SPEACH'
         ];
         this.setAI(
             ...preferences
@@ -1087,13 +1111,14 @@ class AI {
     #speechControlGeneration=0;
     #speechFailureSequence=0;
     #stopOllamaReady=null;
+    #ttsSegmentation={...DEFAULT_TTS_SEGMENTATION};
     #preferenceTuple=completeValue([
         'OPENAI',
+        'LOCAL_SPEACH',
+        'LOCAL_SPEACH',
         'OPENAI',
-        'OPENAI',
-        'OPENAI',
-        'OPENAI',
-        'OPENAI'
+        'LOCAL_SPEACH',
+        'LOCAL_SPEACH'
     ]);
 
     get providerRuntime(){
@@ -1173,9 +1198,6 @@ class AI {
         this.#retainLegacyLLMReadiness(
             this.#reconcileLegacyLLMReadiness()
         );
-        this.#retainLegacySpeechReadiness(
-            this.#reconcileLegacySpeechReadiness()
-        );
         return this.#license;
     }
 
@@ -1427,12 +1449,6 @@ class AI {
         if(role!=='tts'){
             return null;
         }
-        if(providerId==='OPENAI'){
-            const selected=typeof globalThis.window?.user?.AI_voice==='string'
-                ?globalThis.window.user.AI_voice.trim()
-                :'';
-            return selected||'alloy';
-        }
         if(providerId==='LOCAL_SPEACH'){
             return 'af_heart';
         }
@@ -1445,10 +1461,6 @@ class AI {
         if(service!==providerId||!model){
             return false;
         }
-        if(providerId==='OPENAI'){
-            return Boolean(this.#legacyOpenAISpeechKey)
-                &&typeof globalThis.fetch==='function';
-        }
         if(providerId==='LOCAL_SPEACH'){
             return Boolean(this.#nativeSpeech(service,role));
         }
@@ -1685,8 +1697,7 @@ class AI {
     }
 
     #ensureLegacySpeechProvider(role,providerId){
-        if(!['stt','tts'].includes(role)
-            ||!['OPENAI','LOCAL_SPEACH'].includes(providerId)){
+        if(!['stt','tts'].includes(role)||providerId!=='LOCAL_SPEACH'){
             return false;
         }
         if(this.#providerRuntime.hasProvider(role,providerId)){
@@ -1875,11 +1886,7 @@ class AI {
             throw error;
         }
 
-        if(service==='OPENAI'&&(
-            role==='llm'
-                ?Boolean(this.twinKey)
-                :Boolean(this.#legacyOpenAISpeechKey)
-        )){
+        if(service==='OPENAI'&&role==='llm'&&Boolean(this.twinKey)){
             return true;
         }
 
@@ -1947,6 +1954,10 @@ class AI {
             }
             return value;
         });
+        next[1]='LOCAL_SPEACH';
+        next[2]='LOCAL_SPEACH';
+        next[4]='LOCAL_SPEACH';
+        next[5]='LOCAL_SPEACH';
         return completeValue(next);
     }
 
@@ -1961,6 +1972,26 @@ class AI {
         }
     }
 
+    #assertDeviceSpeechConfiguration(configuration){
+        for(const role of ['stt','tts']){
+            for(const routeName of ['default','localOnly']){
+                const selection=configuration?.[role]?.[routeName];
+                if(selection&&selection.localOnly!==true){
+                    const operation=role==='stt'
+                        ?'Speech recognition'
+                        :'Speech synthesis';
+                    const error=new TypeError(
+                        `${operation} supports on-device providers only.`
+                    );
+                    error.code=role==='stt'
+                        ?'AI_STT_DEVICE_ONLY'
+                        :'AI_TTS_DEVICE_ONLY';
+                    throw error;
+                }
+            }
+        }
+    }
+
     #normalizedLLMModel(service,model){
         if(service==='OLLAMA'){
             const mappedModel=model==='OPENAI'?null:this.#models[model];
@@ -2187,6 +2218,7 @@ class AI {
 
     configureProviders(selections){
         const prepared=this.#providerRuntime.validateConfiguration(selections);
+        this.#assertDeviceSpeechConfiguration(prepared);
         this.#assertSynchronousBrowserSpeechSupersession('AI.configureProviders');
         this.#ensureLegacyLLMProvider(
             prepared.llm.default?.providerId
@@ -2221,6 +2253,7 @@ class AI {
 
     configureSpeechProviders(selections){
         const prepared=this.#providerRuntime.validateSpeechConfiguration(selections);
+        this.#assertDeviceSpeechConfiguration(prepared);
         this.#assertSynchronousBrowserSpeechSupersession(
             'AI.configureSpeechProviders'
         );
@@ -2287,6 +2320,7 @@ class AI {
 
     async transitionProviders(selections){
         const prepared=this.#providerRuntime.validateConfiguration(selections);
+        this.#assertDeviceSpeechConfiguration(prepared);
         await this.#supersedeBrowserSpeechForRouteChange();
         this.#ensureLegacyLLMProvider(
             prepared.llm.default?.providerId
@@ -2318,6 +2352,7 @@ class AI {
 
     async transitionSpeechProviders(selections){
         const prepared=this.#providerRuntime.validateSpeechConfiguration(selections);
+        this.#assertDeviceSpeechConfiguration(prepared);
         await this.#supersedeBrowserSpeechForRouteChange();
         this.#invalidateSpeechControl();
         await this.#unloadSpeechProviderRolesForTransition();
@@ -5520,6 +5555,18 @@ class AI {
         return responseJSON;
     }
 
+    get ttsSegmentation(){
+        return {...this.#ttsSegmentation};
+    }
+
+    configureTTSSegmentation(options={}){
+        this.#ttsSegmentation=normalizeTTSSegmentation(
+            options,
+            this.#ttsSegmentation
+        );
+        return this.ttsSegmentation;
+    }
+
     streamTTS(
         text='',
         end=false
@@ -5571,10 +5618,16 @@ class AI {
         let remainder=this.audioMessageChunks;
 
         while(remainder.length>0){
-            const terminator=this.#findSpeechTerminator(remainder,end);
-            let boundary=terminator
-                ?terminator.index+terminator[0].length
-                :-1;
+            const punctuationBoundary=this.#findSpeechPunctuationBoundary(
+                remainder,
+                end
+            );
+            const cadenceBoundary=this.#findSpeechCadenceBoundary(remainder,end);
+            let boundary=punctuationBoundary;
+
+            if(boundary<0||(cadenceBoundary>0&&cadenceBoundary<boundary)){
+                boundary=cadenceBoundary;
+            }
 
             if(boundary<0&&end){
                 boundary=remainder.length;
@@ -5596,7 +5649,14 @@ class AI {
         return segments;
     }
 
-    #findSpeechTerminator(text,end=false){
+    #findSpeechPunctuationBoundary(text,end=false){
+        if(this.#ttsSegmentation.punctuation==='none'){
+            return -1;
+        }
+        if(this.#ttsSegmentation.punctuation==='any'){
+            return this.#findAnySpeechPunctuationBoundary(text,end);
+        }
+
         const pattern=end
             ?/(?:[\u3002\uFF01\uFF1F]|[.!?](?=\s|$))/g
             :/(?:[\u3002\uFF01\uFF1F]|[.!?](?=\s+\S))/g;
@@ -5610,10 +5670,71 @@ class AI {
                 continue;
             }
 
-            return terminator;
+            return this.#includeSpeechBoundaryWhitespace(
+                text,
+                terminator.index+terminator[0].length
+            );
         }
 
-        return null;
+        return -1;
+    }
+
+    #findAnySpeechPunctuationBoundary(text,end=false){
+        const pattern=end
+            ?TTS_ANY_PUNCTUATION_AT_END
+            :TTS_ANY_PUNCTUATION;
+        pattern.lastIndex=0;
+        const firstContent=text.search(/\S/u);
+        let punctuation;
+
+        while((punctuation=pattern.exec(text))){
+            if(firstContent<0||punctuation.index<=firstContent){
+                continue;
+            }
+            const boundary=punctuation.index+punctuation[0].length;
+            return this.#includeSpeechBoundaryWhitespace(text,boundary);
+        }
+
+        return -1;
+    }
+
+    #findSpeechCadenceBoundary(text,end=false){
+        const wordCadence=this.#ttsSegmentation.wordCadence;
+        if(wordCadence===null){
+            return -1;
+        }
+
+        const words=/\S+/gu;
+        let word;
+        let wordCount=0;
+
+        while((word=words.exec(text))){
+            wordCount+=1;
+            if(wordCount!==wordCadence){
+                continue;
+            }
+            const boundary=word.index+word[0].length;
+            if(boundary===text.length&&!end){
+                return -1;
+            }
+            if(boundary<text.length&&!/\s/u.test(text[boundary])){
+                return -1;
+            }
+            return this.#includeSpeechBoundaryWhitespace(text,boundary);
+        }
+
+        return -1;
+    }
+
+    #includeSpeechBoundaryWhitespace(text,boundary){
+        let completeBoundary=boundary;
+        while(
+            completeBoundary<text.length
+            &&/\s/u.test(text[completeBoundary])
+        ){
+            completeBoundary+=1;
+        }
+        return completeBoundary;
     }
 
     #isSpeechAbbreviation(text,periodIndex){