Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Fix emoji reading in chat read-aloud (#16480)

rugved8607346f423d1025
diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index 791ddf042d7..353b3422af1 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -48,6 +48,7 @@
"decode-named-character-reference": "1.3.0",
"dexie": "^4.3.0",
"dingbat-to-unicode": "^1.0.1",
+ "emoji-regex": "11.0.0",
"fflate": "0.8.3",
"js-yaml": "^4.3.2",
"katex": "^0.16.28",
@@ -8249,11 +8250,10 @@
"license": "ISC"
},
"node_modules/emoji-regex": {
- "version": "8.0.0",
- "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
- "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
- "license": "MIT",
- "optional": true
+ "version": "11.0.0",
+ "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-11.0.0.tgz",
+ "integrity": "sha512-2qQtc7syPcJmvF7BfKxBvwgvhzDth6u3q23+Jr4JOQUcmMGgmj7B4/ePZ2DmOhrrb77S8kMNHq2kilVASysfqA==",
+ "license": "MIT"
},
"node_modules/enhanced-resolve": {
"version": "5.21.0",
@@ -12101,6 +12101,13 @@
"node": ">=8"
}
},
+ "node_modules/string-width/node_modules/emoji-regex": {
+ "version": "8.0.0",
+ "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz",
+ "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==",
+ "license": "MIT",
+ "optional": true
+ },
"node_modules/string-width/node_modules/strip-ansi": {
"version": "6.0.1",
"resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz",
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index 62305ddf182..6c3ec82bddf 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -63,6 +63,7 @@
"decode-named-character-reference": "1.3.0",
"dexie": "^4.3.0",
"dingbat-to-unicode": "^1.0.1",
+ "emoji-regex": "11.0.0",
"fflate": "0.8.3",
"js-yaml": "^4.3.2",
"katex": "^0.16.28",
diff --git a/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts
index ebc131c1220..4422eaf1b5e 100644
--- a/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts
+++ b/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts
@@ -119,7 +119,6 @@ function langBase(tag: string): string {
const MAX_CURATED_VOICES = 20;
-/** Filter, rank and cap voices; the selected voice is always kept. */
export function curateSystemVoices(
voices: SpeechSynthesisVoice[],
selectedVoiceURI?: string,
@@ -261,7 +260,6 @@ function readAloudError(
return new Error(detail);
}
-/** Generate speech via the loaded TTS audio model; returns a WAV data URL. */
export async function generateStudioTtsAudio(
text: string,
signal?: AbortSignal,
@@ -421,7 +419,6 @@ function speakWithBackendAudio(
let audioUrl: string | null = null;
let cancelled = false;
- // Release the element and its multi-MB WAV data URL as soon as playback ends.
const cleanup = () => {
if (audio) {
audio.pause();
@@ -477,7 +474,6 @@ function speakWithBackendAudio(
};
}
-/** Engines: "system" (speechSynthesis), "studio" (local TTS), "custom" (a connection). */
export class StudioSpeechSynthesisAdapter implements SpeechSynthesisAdapter {
static systemVoicesSupported(): boolean {
return (
@@ -504,7 +500,6 @@ export class StudioSpeechSynthesisAdapter implements SpeechSynthesisAdapter {
error?: unknown,
) => {
if (res.status.type === "ended") return;
- // A cancelled or interrupted utterance is a normal stop and must not toast.
if (
reason === "error" &&
error !== "interrupted" &&
@@ -540,7 +535,11 @@ export class StudioSpeechSynthesisAdapter implements SpeechSynthesisAdapter {
};
},
};
-
+ if (!text.trim()) {
+ cancelImpl = () => {};
+ handleEnd("finished");
+ return res;
+ }
if (
ttsEngine !== "system" ||
!StudioSpeechSynthesisAdapter.systemVoicesSupported()
diff --git a/studio/frontend/src/features/chat/utils/speech-text.ts b/studio/frontend/src/features/chat/utils/speech-text.ts
index f5a72de712a..595b5deb8e2 100644
--- a/studio/frontend/src/features/chat/utils/speech-text.ts
+++ b/studio/frontend/src/features/chat/utils/speech-text.ts
@@ -2,6 +2,7 @@
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import { decodeNamedCharacterReference } from "decode-named-character-reference";
+import emojiRegex from "emoji-regex";
import type { Nodes, Root, RootContent } from "mdast";
import { fromMarkdown } from "mdast-util-from-markdown";
import { gfmFromMarkdown } from "mdast-util-gfm";
@@ -23,6 +24,27 @@ const LINE_TAGS = new Set(
" ",
),
);
+function removeTtsEmojis(text: string): string {
+ let removed = false;
+ const cleaned = text.replace(emojiRegex(), (emoji, offset, source) => {
+ if (
+ source[offset + emoji.length] === "\uFE0E" ||
+ ([...emoji].length === 1 &&
+ /\p{Emoji}/u.test(emoji) &&
+ !/\p{Emoji_Presentation}/u.test(emoji))
+ ) {
+ return emoji;
+ }
+ removed = true;
+ return " ";
+ });
+ return removed
+ ? cleaned
+ .replace(/[ \t]+([,.;:!?])/g, "$1")
+ .replace(/[ \t]{2,}/g, " ")
+ .trim()
+ : text;
+}
/** The words a markdown reply shows, for read-aloud: voices speak raw markup ("asterisk asterisk", #12547). */
export function markdownToSpeechText(markdown: string): string {
@@ -34,7 +56,7 @@ export function markdownToSpeechText(markdown: string): string {
mdastExtensions: [gfmFromMarkdown(), mathFromMarkdown()],
});
} catch {
- return markdown;
+ return removeTtsEmojis(markdown);
}
const out: string[] = [];
const notes: string[] = [];
@@ -46,7 +68,9 @@ export function markdownToSpeechText(markdown: string): string {
false,
);
}
- return [...out, ...notes].filter((block) => block.length > 0).join("\n");
+ return removeTtsEmojis(
+ [...out, ...notes].filter((block) => block.length > 0).join("\n"),
+ );
}
function collectBlocks(
@@ -68,7 +92,10 @@ function collectBlocks(
return;
case "table":
for (const row of node.children) {
- out.push(sentence(row.children.map(inline).join(", ")));
+ const cells = row.children.map((cell) => removeTtsEmojis(inline(cell)));
+ if (cells.some((cell) => cell.trim())) {
+ out.push(sentence(cells.join(", ")));
+ }
}
return;
// The sanitizer unwraps raw HTML, so its text shows on the page.
@@ -138,7 +165,7 @@ function htmlText(html: string): string {
}
function sentence(text: string): string {
- const trimmed = text.trim();
+ const trimmed = removeTtsEmojis(text).trim();
if (!trimmed || /[.!?:;,…]$/.test(trimmed)) return trimmed;
return `${trimmed}.`;
}
diff --git a/studio/frontend/tests/speech-text.test.ts b/studio/frontend/tests/speech-text.test.ts
index 9e392f9c4e1..ea8bee4ef48 100644
--- a/studio/frontend/tests/speech-text.test.ts
+++ b/studio/frontend/tests/speech-text.test.ts
@@ -114,3 +114,61 @@ test("HTML tags keep the word boundaries the page shows", () => {
);
assert.equal(markdownToSpeechText("<div>H<sub>2</sub>O</div>"), "H2O");
});
+
+test("read-aloud skips emoji sequences but preserves surrounding words", () => {
+ const spoken = markdownToSpeechText(
+ "Japan 🇯🇵 developer 👩‍💻 thumbs 👍🏽 heart ❤️",
+ );
+
+ assert.match(spoken, /Japan\s+developer\s+thumbs\s+heart/);
+
+ for (const emoji of ["🇯🇵", "👩‍💻", "👍🏽", "❤️"]) {
+ assert.equal(spoken.includes(emoji), false);
+ }
+});
+
+test("read-aloud handles emoji-only messages", () => {
+ assert.equal(markdownToSpeechText("😀 👨‍💻 🇯🇵").trim(), "");
+ for (const input of [
+ "# 😀",
+ "- 😀\n- 👍",
+ "1. 😀",
+ "| 😀 | 😘 |\n| --- | --- |\n| 🥰 | ❤️ |",
+ ]) {
+ assert.equal(markdownToSpeechText(input), "", input);
+ }
+});
+
+test("read-aloud preserves word boundaries around emojis", () => {
+ for (const input of [
+ "Hello😀world",
+ "Hello😀😀world",
+ "Hello👩‍💻👍🏽world",
+ ]) {
+ assert.equal(markdownToSpeechText(input), "Hello world", input);
+ }
+});
+
+test("read-aloud cleans the reported German and English replies", () => {
+ assert.equal(
+ markdownToSpeechText("Hallo! 🥰❤️ Schön, dass du da bist. 😘"),
+ "Hallo! Schön, dass du da bist.",
+ );
+ assert.equal(
+ markdownToSpeechText("Hello! 🥰❤️ It's wonderful to see you. 😘"),
+ "Hello! It's wonderful to see you.",
+ );
+});
+
+test("read-aloud preserves ordinary symbols and whitespace without emojis", () => {
+ assert.equal(
+ markdownToSpeechText("© 2026 ® ™ ℹ 123 # *"),
+ "© 2026 ® ™ ℹ 123 # *",
+ );
+ assert.equal(markdownToSpeechText("`two spaces`"), "two spaces");
+ assert.equal(markdownToSpeechText("word🥰❤️, next"), "word, next");
+ assert.equal(markdownToSpeechText("$A ↔ B$"), "A ↔ B");
+ assert.equal(markdownToSpeechText("Sex: ♀ / ♂"), "Sex: ♀ / ♂");
+ assert.equal(markdownToSpeechText("Watch ⌚\uFE0E"), "Watch ⌚\uFE0E");
+ assert.equal(markdownToSpeechText("©️ ®️ ™️ ℹ️"), "");
+});