diff --git a/backend/internal/cfai/client.go b/backend/internal/cfai/client.go
index 53347ec..078432a 100644
--- a/backend/internal/cfai/client.go
+++ b/backend/internal/cfai/client.go
@@ -17,6 +17,10 @@
 // response. There is no 100-second Cloudflare proxy timeout because we are
 // calling the Cloudflare API directly, not routing through a Cloudflare-proxied
 // homelab tunnel.
+//
+// The aura-2-en model enforces a hard 2 000-character limit per request.
+// GenerateAudio transparently splits longer texts into sentence-boundary chunks
+// and concatenates the resulting MP3 frames.
 package cfai
 
 import (
@@ -145,6 +149,8 @@ func New(accountID, apiToken, model string) Client {
 }
 
 // GenerateAudio calls the Cloudflare Workers AI TTS endpoint and returns MP3 bytes.
+// The aura-2-en model rejects inputs longer than 2 000 characters, so this method
+// splits the text into sentence-bounded chunks and concatenates the MP3 responses.
 func (c *httpClient) GenerateAudio(ctx context.Context, text, voice string) ([]byte, error) {
 	if text == "" {
 		return nil, fmt.Errorf("cfai: empty text")
@@ -154,6 +160,20 @@ func (c *httpClient) GenerateAudio(ctx context.Context, text, voice string) ([]b
 		speaker = "luna"
 	}
 
+	chunks := splitText(text, 1800) // stay comfortably under the 2 000-char limit
+	var combined []byte
+	for _, chunk := range chunks {
+		part, err := c.generateChunk(ctx, chunk, speaker)
+		if err != nil {
+			return nil, err
+		}
+		combined = append(combined, part...)
+	}
+	return combined, nil
+}
+
+// generateChunk sends a single ≤2 000-character request and returns MP3 bytes.
+func (c *httpClient) generateChunk(ctx context.Context, text, speaker string) ([]byte, error) {
 	body, err := json.Marshal(map[string]any{
 		"text":    text,
 		"speaker": speaker,
@@ -189,6 +209,87 @@ func (c *httpClient) GenerateAudio(ctx context.Context, text, voice string) ([]b
 	return mp3, nil
 }
 
+// splitText splits src into chunks of at most maxChars characters each.
+// It tries to break at paragraph boundaries first, then at sentence-ending
+// punctuation (. ! ?), and falls back to the nearest space.
+func splitText(src string, maxChars int) []string {
+	if len(src) <= maxChars {
+		return []string{src}
+	}
+
+	var chunks []string
+	remaining := src
+
+	for len(remaining) > 0 {
+		if len(remaining) <= maxChars {
+			chunks = append(chunks, strings.TrimSpace(remaining))
+			break
+		}
+
+		// Search window: the first maxChars bytes of remaining.
+		// Use byte length here because the API limit is in bytes/chars for ASCII;
+		// for safety we operate on rune-aware slices.
+		window := remaining
+		if len(window) > maxChars {
+			// Trim to maxChars runes (not bytes), ensuring we don't split a multi-byte char.
+			window = runeSlice(remaining, maxChars)
+		}
+
+		cut := -1
+
+		// 1. Prefer paragraph break (\n\n or \n).
+		if i := strings.LastIndex(window, "\n\n"); i > 0 {
+			cut = i + 2
+		} else if i := strings.LastIndex(window, "\n"); i > 0 {
+			cut = i + 1
+		}
+
+		// 2. Fall back to sentence-ending punctuation followed by a space.
+		if cut < 0 {
+			for _, punct := range []string{". ", "! ", "? ", ".\n", "!\n", "?\n"} {
+				if i := strings.LastIndex(window, punct); i > 0 {
+					candidate := i + len(punct)
+					if cut < 0 || candidate > cut {
+						cut = candidate
+					}
+				}
+			}
+		}
+
+		// 3. Last resort: nearest space.
+		if cut < 0 {
+			if i := strings.LastIndex(window, " "); i > 0 {
+				cut = i + 1
+			}
+		}
+
+		// 4. Hard cut at maxChars runes if no boundary found.
+		if cut < 0 {
+			cut = len(window)
+		}
+
+		chunk := strings.TrimSpace(remaining[:cut])
+		if chunk != "" {
+			chunks = append(chunks, chunk)
+		}
+		remaining = remaining[cut:]
+	}
+
+	return chunks
+}
+
+// runeSlice returns the first n runes of s as a string.
+func runeSlice(s string, n int) string {
+	count := 0
+	for i := range s {
+		if count == n {
+			return s[:i]
+		}
+		count++
+	}
+	return s
+}
+
 // StreamAudioMP3 generates audio and wraps the MP3 bytes as an io.ReadCloser.
 func (c *httpClient) StreamAudioMP3(ctx context.Context, text, voice string) (io.ReadCloser, error) {
 	mp3, err := c.GenerateAudio(ctx, text, voice)