diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
index a956f83..f16466c 100644
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@@ -302,7 +302,9 @@ covered by the `styles.css` manifest.
   line editable in place, where else it is used, Main / Aux for the voice picked), with filters,
   a search and the first 60 rows then more. It opens at once and loads the character's clips,
   their lines and lengths (`/anomalous_tts/reference_lines`) in the background (save waits for
-  them), and keeps settings fields it does not know. `tts_voice_refs.js` holds its rules without
+  them), and keeps settings fields it does not know; 🗑 on a row moves that clip to the
+  Recycle Bin through the node (`removeGptSovitsClips`, which also takes it out of the saved
+  settings) and out of the window's voices. `tts_voice_refs.js` holds its rules without
   DOM: the voices read from the settings and the node's view (an emotion found from a file
   name is written only once changed), one main reference and any number of auxiliary ones per
   voice, the lines fixed here (`lines`; a voice's own `text` moves there), what keeps it from
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 78ee37f..ee8b859 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -65,8 +65,9 @@
 - **Reference audio and emotions** (the character card's button, formerly *Edit emotions*) is now a large window. On top, the **voices**: the main voice, each emotion with how many references it has (*1 main, 2 aux*), and **+ New emotion**. Below, **every clip** of the character, one row each: ▶ to hear it, its file and length, its line, and where else it is used (*Main voice · aux*). For the voice picked, a row is its **Main** reference or one of its **Aux** references, or nothing; unused clips just stay in the list. Filters (*All*, *This voice*, *In use*, *No line*) and a search by line or file name; the first 60 rows, then more.
 - **Auxiliary references (副参考音频)**: a voice keeps one main reference (its tone and way of speaking, 3–10 s, with a line) and can take any number of auxiliary ones, whose timbre is averaged in when it speaks, as GPT-SoVITS does with several references; they need no line. Only v1 / v2 / v2Pro models use them (the controls say so for v3 / v4).
 - **A clip's line can be fixed right in its row** (改参考台词): the fixed line is kept in the character's settings and used before its .txt / .list files, which stay as they are. A line that is missing shows *No line found: click to type it*.
+- **🗑 on a row moves that clip to the Recycle Bin** (删除参考音频), with its own .txt / .lab, after saying where it is used; it is taken out of those voices and the settings. Never a permanent delete: where there is no Recycle Bin (a USB stick, a network drive) nothing is deleted and the window says so.
 - Clips too long or too short to be a main reference are marked (they can still be auxiliary). The main voice can go back to the node's own pick; an emotion can be renamed or deleted (one found from a file name is changed by renaming the file). Closing with unsaved changes asks first.
-- Auxiliary references, fixed lines and lengths need Anomalous_TTS interface 18 (1.1.0); with an older version the window works as before for main references and says what needs the update.
+- Auxiliary references, fixed lines, lengths and 🗑 need Anomalous_TTS interface 18 (1.1.0); with an older version the window works as before for main references and says what needs the update.
 
 ### 🔍 Scanning (扫描)
 - **The scan page opens at once, also with many models**: it shows the last counts right away, says it is updating them, and shows the fresh ones when they come. What each model file's header says is now kept on disk (`user/anomalous/model_heads.json`, by file, read again only when the file changes), so a big library or a network drive is not read again after every restart either; requests that come while the page counts share that one count.
diff --git a/docs/guides/audio-studio.md b/docs/guides/audio-studio.md
index de7559c..85d23ce 100644
--- a/docs/guides/audio-studio.md
+++ b/docs/guides/audio-studio.md
@@ -214,9 +214,10 @@
 3. **主参考**：决定这个声音的语气和念法。每个声音一条，要 3~10 秒、有台词（太长太短的会标黄，不能当主参考）。主声音也可以 **改回自动挑**，让节点自己选。
 4. **副参考**：只用来融合音色，让声音更稳、更像本人。可以勾好几条，不需要台词，建议同一个人、音质干净的 2~4 条。只有 v1 / v2 / v2Pro 模型用得上，v3 / v4 的角色这一列是灰的。
 5. **改台词**：直接点一行的台词就能改；找不到台词的会写“没找到台词，点这里填”。改过的台词存在角色的设置里，比 `.txt` / 标注文件优先，原来的文件不动。清空就换回文件里的台词。
-6. 点 **保存**。设置会写进角色文件夹里的 `anomalous_tts.json`，由节点检查后保存。有问题（比如某个情绪还没选主参考）会提示，什么都不会改。没保存就关窗口会先问一句。
+6. **删除**：点一行最右边的 🗑，确认后这条音频和它同名的 `.txt` / `.lab` 一起**移到系统回收站**（随时能还原，从不永久删除；U 盘、网络盘上没有回收站就不删，会说明）。它用在哪些声音里，确认框里会写，删掉后自动从这些声音和设置里去掉。
+7. 点 **保存**。设置会写进角色文件夹里的 `anomalous_tts.json`，由节点检查后保存。有问题（比如某个情绪还没选主参考）会提示，什么都不会改。没保存就关窗口会先问一句。
 
-副参考、改台词和音频长度需要 Anomalous_TTS 接口版本 18（1.1.0）；旧版本只能选主参考，窗口里会说明。
+副参考、改台词、音频长度和删除需要 Anomalous_TTS 接口版本 18（1.1.0）；旧版本只能选主参考，窗口里会说明。
 
 文件名本身带情绪的音频（例如 `Arona_Talk.开心.wav`）会被自动识别，不用再填一遍。只是扩展名的部分不算情绪，例如 `Arona_Talk.ogg.wav`、`Arona_Talk.ogg (1).ogg`。
 
diff --git a/web/modules/audio_engines.js b/web/modules/audio_engines.js
index 38c53b7..b9404e4 100644
--- a/web/modules/audio_engines.js
+++ b/web/modules/audio_engines.js
@@ -208,6 +208,25 @@ export async function fetchGptSovitsDetail(name, signal) {
 
 // ---------- GPT-SoVITS settings (Anomalous writes, the node reads) ----------
 
+/**
+ * Move a character's reference clips (and their own .txt / .lab) to the Recycle Bin through
+ * the node (interface 18); it also takes them out of the settings. Returns the node's answer
+ * (`removed`, `text_files`, `character`: the detail after).
+ */
+export async function removeGptSovitsClips(name, paths) {
+    const resp = await fetch('/anomalous_tts/audio/remove', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ character: name, paths }),
+    });
+    if (!resp.ok) {
+        const message = await resp.text().catch(() => '');
+        throw new Error(message || `HTTP ${resp.status}`);
+    }
+    invalidateEngineCache();
+    return resp.json();
+}
+
 /**
  * Replace a character's anomalous_tts.json through the node's API. `settings` must be the
  * full object (unknown fields preserved by the caller). Returns the refreshed group.
diff --git a/web/modules/locales.js b/web/modules/locales.js
index e9ce3d3..76946b0 100644
--- a/web/modules/locales.js
+++ b/web/modules/locales.js
@@ -2154,6 +2154,11 @@ export const i18n = {
         ttsRefsShowMore: "再显示一些（还有 {count} 条）",
         ttsRefsNoMain: "情绪“{name}”还没选主参考。",
         ttsRefsCloseUnsaved: "有改动还没保存，确定关闭吗？",
+        ttsRefsRemove: "移到回收站",
+        ttsRefsRemoveConfirm: "把 {file} 移到回收站？它同名的台词文件（.txt / .lab）也会一起移走，需要时可以从回收站还原。",
+        ttsRefsRemoveUses: "它现在是 {uses} 的参考，会从这些声音里去掉。",
+        ttsRefsRemoveFailed: "没有删除：{error}",
+        ttsRefsRemoved: "已把 {file} 移到回收站。这个角色还有 {count} 条音频。",
         ttsEditorRemove: "删除这一行",
         ttsEditorSave: "保存",
         ttsEditorSaveFailed: "保存失败：{error}",
@@ -4846,6 +4851,11 @@ export const i18n = {
         ttsRefsShowMore: "Show more ({count} left)",
         ttsRefsNoMain: "The emotion “{name}” has no main reference yet.",
         ttsRefsCloseUnsaved: "There are unsaved changes. Close anyway?",
+        ttsRefsRemove: "Move to the Recycle Bin",
+        ttsRefsRemoveConfirm: "Move {file} to the Recycle Bin? Its own line file (.txt / .lab) goes with it; both can be restored from the Recycle Bin.",
+        ttsRefsRemoveUses: "It is now a reference of {uses}, and is taken out of those voices.",
+        ttsRefsRemoveFailed: "Nothing was deleted: {error}",
+        ttsRefsRemoved: "{file} is in the Recycle Bin. {count} audio files left in this character.",
         ttsEditorRemove: "Remove this row",
         ttsEditorSave: "Save",
         ttsEditorSaveFailed: "Could not save: {error}",
diff --git a/web/modules/ui_audio_tts_editor.js b/web/modules/ui_audio_tts_editor.js
index 7c042af..8384e0a 100644
--- a/web/modules/ui_audio_tts_editor.js
+++ b/web/modules/ui_audio_tts_editor.js
@@ -1,7 +1,7 @@
 import { t } from './interface_settings.js';
 import { createViewScope } from './ui_lifecycle.js';
 import { anomalousAlert, anomalousConfirm, anomalousPrompt } from './ui_dialog.js';
-import { fetchGptSovitsDetail, fetchReferenceClips, saveGptSovitsSettings, ttsAudioUrl } from './audio_engines.js';
+import { fetchGptSovitsDetail, fetchReferenceClips, gptSovitsGroups, removeGptSovitsClips, saveGptSovitsSettings, ttsAudioUrl } from './audio_engines.js';
 import {
     AUX_FORMAT, AUX_VERSIONS, LINE_LIMIT, MAIN, clearMain, fitsMain, listClips, nameProblem, readVoices, roleOf,
     setMain, toggleAux, usesOf, voiceCount, voicesProblem, writeSettings,
@@ -15,8 +15,10 @@ import {
  * **auxiliary references** (their timbre is averaged in; any number, no line needed). A clip's
  * line can be fixed right in its row. The rules are in tts_voice_refs.js; the node owns the
  * file, this window only sends the new settings through its API, keeping every field it does
- * not edit. Auxiliary references and fixed lines need Anomalous_TTS interface 18 (and
- * auxiliary references a v1 / v2 / v2Pro model); otherwise those controls are off and say why.
+ * not edit. A clip can be moved to the Recycle Bin from its row (the node moves it with its own
+ * .txt / .lab and takes it out of the settings). Auxiliary references, fixed lines and removing
+ * need Anomalous_TTS interface 18 (and auxiliary references a v1 / v2 / v2Pro model); otherwise
+ * those controls are off or hidden and say why.
  *
  * The window opens at once and reads the character's clips, their lines and lengths in the
  * background (a folder can hold a thousand clips: the list shows the first ones, then more).
@@ -320,10 +322,52 @@ export function openGptSovitsEditor(group, { onSaved } = {}) {
         auxBtn.disabled = !current || !canAux() || role === 'main';
         if (!canAux()) auxBtn.title = auxReason();
         roles.append(mainBtn, auxBtn);
+        if (canFix()) {
+            roles.append(button('anomalous-tts-refs-remove', '🗑', () => removeClip(path), t('ttsRefsRemove')));
+        }
         row.append(playBtn, file, middle, roles);
         return row;
     }
 
+    /** One clip to the Recycle Bin (after asking), then out of this window's voices too. */
+    async function removeClip(path) {
+        const uses = usesOf(state.voices, path).map(use => t(`ttsRefsUse_${use.role}`, { name: use.name === MAIN ? t('ttsRefsMainVoice') : use.name }));
+        const message = [
+            t('ttsRefsRemoveConfirm', { file: baseName(path) }),
+            uses.length ? t('ttsRefsRemoveUses', { uses: uses.join('、') }) : '',
+        ].filter(Boolean).join('\n');
+        if (!await anomalousConfirm(message, t('ttsRefsRemove'))) return;
+        if (state.playing?.path === path) stopPlaying();
+        let answer;
+        try {
+            answer = await removeGptSovitsClips(name, [path]);
+        } catch (e) {
+            if (!scope.signal.aborted) await anomalousAlert(t('ttsRefsRemoveFailed', { error: e.message }));
+            return;
+        }
+        if (scope.signal.aborted) return;
+        state.audio = state.audio.filter(item => item !== path);
+        state.found.delete(path);
+        state.seconds.delete(path);
+        state.fixes.delete(path);
+        state.cleared.delete(path);
+        // The node took it out of the saved settings; an emotion it was the main reference of went
+        // there, and goes here too unless it was changed in this window (then it waits for another).
+        state.voices = state.voices.filter(item => !(item.main === path && item.name !== MAIN && !item.touched));
+        for (const item of state.voices) {
+            item.aux = item.aux.filter(other => other !== path);
+            if (item.main === path) item.main = null;
+        }
+        if (!state.voices.some(item => item.key === state.voiceKey)) state.voiceKey = 'main';
+        const after = answer?.character;
+        if (after?.settings && typeof after.settings === 'object') state.settings = after.settings;
+        if (after) raw.reference = after.reference;
+        status.textContent = t('ttsRefsRemoved', { file: baseName(path), count: state.audio.length });
+        order = order.filter(item => item !== path);
+        renderAll(false);
+        onSaved?.(after ? gptSovitsGroups({ characters: [after] })[0] : null);
+    }
+
     function renderRow(path) {
         const old = rows.get(path);
         if (!old) return;
diff --git a/web/styles.css b/web/styles.css
index c94ddaa..e74946a 100644
--- a/web/styles.css
+++ b/web/styles.css
@@ -34,4 +34,4 @@
 @import url("./styles/33-kept-combos.css?v=20261019-hero");
 @import url("./styles/34-voice-over-panel.css?v=20261020-layers");
 @import url("./styles/35-voice-over-drafts.css?v=20261020-drafts");
-@import url("./styles/36-reference-audio.css?v=20261020-refs");
+@import url("./styles/36-reference-audio.css?v=20261020-remove");
diff --git a/web/styles/36-reference-audio.css b/web/styles/36-reference-audio.css
index 31b470e..d6d073a 100644
--- a/web/styles/36-reference-audio.css
+++ b/web/styles/36-reference-audio.css
@@ -279,6 +279,29 @@
     cursor: not-allowed;
 }
 
+/* To the Recycle Bin: quiet until pointed at. */
+.anomalous-tts-refs-remove {
+    width: 30px;
+    padding: 4px 0;
+    border: 1px solid transparent;
+    border-radius: 6px;
+    background: none;
+    color: var(--amb-text-dim);
+    font-size: 13px;
+    cursor: pointer;
+    opacity: 0.55;
+}
+
+.anomalous-tts-refs-row:hover .anomalous-tts-refs-remove {
+    opacity: 1;
+}
+
+.anomalous-tts-refs-remove:hover {
+    border-color: color-mix(in srgb, var(--amb-danger) 50%, transparent);
+    background: color-mix(in srgb, var(--amb-danger) 14%, transparent);
+    color: var(--amb-danger-text);
+}
+
 .anomalous-tts-refs-more,
 .anomalous-tts-refs-empty {
     display: block;