speechready 0.1.0 → 0.2.0

package/README.md

@@ -57,6 +57,17 @@ for (const token of llmStream) {
 for (const sentence of seg.flush()) speak(sentence);
 ```
 
+For LiveKit Agents (Node), the same thing as a `ttsTextTransforms` entry; sentences are released as they are confirmed, and a fenced code block is handled as one piece:
+
+```js
+import { textTransform } from "speechready/livekit";
+
+const session = new voice.AgentSession({
+  ...,
+  ttsTextTransforms: [textTransform({ provider: "elevenlabs", model: "eleven_turbo_v2_5" })],
+});
+```
+
 Pronunciation lexicon, one file for every vendor:
 
 ```js

package/dist/chunk.d.ts

@@ -11,16 +11,12 @@
  */
 /** Split text into sentences. Newlines are hard boundaries. Empty sentences are dropped. */
 export declare function splitSentences(text: string, locale?: string): string[];
-/** Streaming sentence segmenter for LLM token feeds.
- *
- * `push(text)` returns the sentences that are now complete; `flush()` returns
- * whatever is left. A boundary is only confirmed once text after it arrives,
- * so a sentence is never cut at "Dr." or "3.14" because the stream paused there.
- */
 export declare class Segmenter {
     locale: string;
     maxBuffer: number | null;
     private buffer;
+    /** Lines of an open fenced code block; released whole once the fence closes. */
+    private fence;
     constructor(init?: {
         locale?: string;
         maxBuffer?: number | null;

package/dist/chunk.js

@@ -93,10 +93,13 @@ export function splitSentences(text, locale = "en-US") {
  * whatever is left. A boundary is only confirmed once text after it arrives,
  * so a sentence is never cut at "Dr." or "3.14" because the stream paused there.
  */
+const FENCE = /^\s*(```|~~~)/u;
 export class Segmenter {
     locale;
     maxBuffer;
     buffer = "";
+    /** Lines of an open fenced code block; released whole once the fence closes. */
+    fence = null;
     constructor(init = {}) {
         this.locale = init.locale ?? "en-US";
         this.maxBuffer = init.maxBuffer ?? null;
@@ -110,8 +113,22 @@ export class Segmenter {
                 break;
             const head = this.buffer.slice(0, nl);
             this.buffer = this.buffer.slice(nl + 1);
-            done.push(...splitSentences(head, this.locale));
+            if (this.fence !== null) {
+                this.fence.push(head);
+                if (FENCE.test(head)) {
+                    done.push(this.fence.join("\n"));
+                    this.fence = null;
+                }
+            }
+            else if (FENCE.test(head)) {
+                this.fence = [head];
+            }
+            else {
+                done.push(...splitSentences(head, this.locale));
+            }
         }
+        if (this.fence !== null)
+            return done.filter((s) => s); // inside a fence: nothing is sentence-split
         const sentences = splitSentences(this.buffer, this.locale);
         if (sentences.length > 1) {
             done.push(...sentences.slice(0, -1));
@@ -131,6 +148,11 @@ export class Segmenter {
     flush() {
         const rest = this.buffer;
         this.buffer = "";
+        if (this.fence !== null) {
+            const block = [...this.fence, ...(rest ? [rest] : [])].join("\n");
+            this.fence = null;
+            return [block];
+        }
         return splitSentences(rest, this.locale);
     }
 }

package/dist/integrations/livekit.d.ts

@@ -0,0 +1,19 @@
+/** speechready as a LiveKit Agents text transform.
+ *
+ * LiveKit Agents for Node passes the LLM's text stream through `ttsTextTransforms`
+ * before synthesis. Each transform is `(AsyncIterable<string>) => AsyncIterable<string>`.
+ * This one segments the stream into sentences as they are confirmed, runs each
+ * through `prepare`, and yields the result.
+ *
+ *     import { textTransform } from "speechready/livekit";
+ *     const session = new voice.AgentSession({
+ *       ...,
+ *       ttsTextTransforms: [textTransform({ provider: "elevenlabs", model: "eleven_turbo_v2_5" })],
+ *     });
+ */
+import type { PrepareInit } from "../pipeline.js";
+export interface TextTransformInit extends Omit<PrepareInit, "limit"> {
+    /** Release a long sentence at a clause boundary once the buffer holds this many characters. */
+    maxBuffer?: number | null;
+}
+export declare function textTransform(init?: TextTransformInit): (text: AsyncIterable<string>) => AsyncIterable<string>;

package/dist/integrations/livekit.js

@@ -0,0 +1,35 @@
+/** speechready as a LiveKit Agents text transform.
+ *
+ * LiveKit Agents for Node passes the LLM's text stream through `ttsTextTransforms`
+ * before synthesis. Each transform is `(AsyncIterable<string>) => AsyncIterable<string>`.
+ * This one segments the stream into sentences as they are confirmed, runs each
+ * through `prepare`, and yields the result.
+ *
+ *     import { textTransform } from "speechready/livekit";
+ *     const session = new voice.AgentSession({
+ *       ...,
+ *       ttsTextTransforms: [textTransform({ provider: "elevenlabs", model: "eleven_turbo_v2_5" })],
+ *     });
+ */
+import { Segmenter } from "../chunk.js";
+import { prepare } from "../pipeline.js";
+export function textTransform(init = {}) {
+    const { maxBuffer = null, ...prepareInit } = init;
+    const locale = prepareInit.locale ?? "en-US";
+    const spoken = (piece) => prepare(piece, prepareInit).text;
+    return async function* transform(text) {
+        const seg = new Segmenter({ locale, maxBuffer });
+        for await (const chunk of text) {
+            for (const piece of seg.push(chunk)) {
+                const out = spoken(piece);
+                if (out)
+                    yield out + " ";
+            }
+        }
+        for (const piece of seg.flush()) {
+            const out = spoken(piece);
+            if (out)
+                yield out + " ";
+        }
+    };
+}

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const version = "0.1.0";
+export declare const version = "0.2.0";

package/dist/version.js

@@ -1,2 +1,2 @@
 // Generated by scripts/build-profiles.mjs from package.json. Do not edit.
-export const version = "0.1.0";
+export const version = "0.2.0";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "speechready",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Prepare text for text-to-speech: clean LLM output, verbalize ambiguous tokens, compile pauses to each provider's syntax, chunk under model limits.",
   "type": "module",
   "main": "./dist/index.js",
@@ -9,18 +9,38 @@
     ".": {
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
+    },
+    "./livekit": {
+      "types": "./dist/integrations/livekit.d.ts",
+      "import": "./dist/integrations/livekit.js"
     }
   },
   "license": "MIT",
   "author": "Berker Batur",
-  "repository": { "type": "git", "url": "https://github.com/berkerbatur/speechready", "directory": "js" },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/berkerbatur/speechready.git",
+    "directory": "js"
+  },
   "homepage": "https://github.com/berkerbatur/speechready",
-  "keywords": ["tts", "text-to-speech", "text-normalization", "ssml", "voice-agents"],
-  "files": ["dist", "README.md", "LICENSE"],
-  "engines": { "node": ">=20" },
+  "keywords": [
+    "tts",
+    "text-to-speech",
+    "text-normalization",
+    "ssml",
+    "voice-agents"
+  ],
+  "files": [
+    "dist",
+    "README.md",
+    "LICENSE"
+  ],
+  "engines": {
+    "node": ">=20"
+  },
   "scripts": {
     "build": "node scripts/build-profiles.mjs && tsc -p tsconfig.json",
-    "test": "npm run build && node --test test/spec.test.mjs test/package.test.mjs",
+    "test": "npm run build && node --test test/spec.test.mjs test/package.test.mjs test/livekit.test.mjs",
     "prepack": "npm run build"
   },
   "devDependencies": {