diff --git a/speech.py b/speech.py
index 1dff98c..7878746 100755
--- a/speech.py
+++ b/speech.py
@@ -1295,39 +1295,33 @@ async def generate_speech(request: GenerateSpeechRequest):
                     logger.info(f"Loading F5-TTS model on device '{device}'")
                     f5_model = await asyncio.to_thread(f5_wrapper, device=device)
 
-        # Split text into sentences for streaming (first audio arrives faster)
-        sentences = simple_sentence_split(input_text, max_length=500)
-        logger.info(f"Split text into {len(sentences)} sentences for F5-TTS streaming")
+        # F5-TTS handles long text via internal chunking. Per-sentence splitting
+        # produces a ref-to-gen transition artifact at the start of EACH chunk
+        # (audible as gibberish/tinny prefix). Pass full text as one inference.
+        logger.info(f"F5-TTS processing {len(input_text)} chars as single inference")
 
         # F5-TTS outputs float32 PCM at 24kHz
         ffmpeg_args = build_ffmpeg_args(response_format, input_format="f32le", sample_rate="24000")
-
-        # F5-TTS accepts speed natively; only apply atempo as a fallback when ffmpeg can do it cheaper
-        # (we pass speed to the model below, leaving ffmpeg speed at 1.0)
-
         ffmpeg_args.extend(["-"])
         ffmpeg_proc = subprocess.Popen(ffmpeg_args, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
 
         in_q = queue.Queue()
 
         def generator():
-            """Process sentences sequentially and feed to queue"""
+            """Run F5-TTS once for the full input text"""
             try:
-                for idx, sentence in enumerate(sentences):
-                    logger.debug(f"F5-TTS processing sentence {idx+1}/{len(sentences)}: {len(sentence)} chars")
-                    audio_bytes = f5_model.tts(
-                        text=sentence,
-                        ref_audio=ref_audio,
-                        ref_text=ref_text,
-                        speed=speed,
-                    )
-                    in_q.put(audio_bytes)
-                    logger.debug(f"F5-TTS: queued sentence {idx+1}/{len(sentences)}")
+                audio_bytes = f5_model.tts(
+                    text=input_text,
+                    ref_audio=ref_audio,
+                    ref_text=ref_text,
+                    speed=speed,
+                )
+                in_q.put(audio_bytes)
             except Exception as e:
-                logger.error(f"F5-TTS streaming error: {e}")
+                logger.error(f"F5-TTS generation error: {e}")
             finally:
                 in_q.put(None)
-                logger.info(f"F5-TTS streaming complete: {len(sentences)} sentences processed")
+                logger.info("F5-TTS generation complete")
 
         def out_writer():
             """Write audio from queue to ffmpeg stdin"""