README.md

@@ -154,7 +154,8 @@ is built whenever that's set — independent of `LLM_BACKEND`). Endpoints:
 - `POST /tts/speech` — body `{ text, voice?, format?, exaggeration?, cfg_weight?,
   temperature? }`; returns `{ audio_base64, format }`. Input is cleaned
   server-side (markdown + emoji stripped) and the generation knobs are clamped
-  to Chatterbox's ranges.
+  to Chatterbox's ranges. Synthesis is serialized (one at a time — the upstream
+  has no GPU lock of its own); a concurrent request gets a fast `429`.
 - `GET /tts/voices` — list the voice library.
 - `POST /tts/voices/upload` — multipart `voice_name` + `voice_file`; clone a
   voice from an uploaded clip (≤25 MB).

src/ai/tts.rs

@@ -18,6 +18,7 @@ use serde::{Deserialize, Serialize};
 use serde_json::json;
 use std::path::Path;
 use std::sync::LazyLock;
+use tokio::sync::Semaphore;
 
 use crate::data::Claims;
 use crate::file_types::{is_audio_file, is_video_file};
@@ -31,6 +32,14 @@ use crate::state::AppState;
 /// upload can't balloon ImageApi memory before we ever forward it.
 const MAX_VOICE_UPLOAD_BYTES: usize = 25 * 1024 * 1024; // 25 MB
 
+/// Serialize speech synthesis: the Chatterbox server has no internal lock or
+/// queue, so concurrent requests contend on the single GPU and cascade into
+/// timeouts. One permit; when busy we fast-fail with 429 rather than queue —
+/// the app surfaces "busy" immediately, and typical jobs clear in well under a
+/// minute. (An abandoned upstream job can still occupy the GPU until it
+/// finishes — that's a wrapper limitation; the chunked-queue plan fixes it.)
+static TTS_PERMIT: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(1));
+
 /// Sanitize a user-supplied voice name. The name is forwarded to Chatterbox
 /// where it becomes a filename in the voice-library directory, so we restrict
 /// it to a safe charset (alphanumerics, dash, underscore) — no path
@@ -235,6 +244,14 @@ pub async fn tts_speech_handler(
     let cfg_weight = req.cfg_weight.map(|x| x.clamp(0.0, 1.0));
     let temperature = req.temperature.map(|x| x.clamp(0.05, 5.0));
 
+    // One synthesis at a time (see TTS_PERMIT) — fast-fail when busy.
+    let Ok(_permit) = TTS_PERMIT.try_acquire() else {
+        span.set_status(Status::error("tts busy"));
+        return HttpResponse::TooManyRequests().json(json!({
+            "error": "TTS is busy with another request — try again shortly"
+        }));
+    };
+
     match client
         .text_to_speech(&text, voice, format, exaggeration, cfg_weight, temperature)
         .await