It was running great. The problem started after I upgraded my unRAID server where this is running via Docker. My guess is something downstream changed but I have been unable to figure it out in a few hours of troubleshooting. Maybe someone here has seen this before?
There are long pauses after each chunk (each progress bar in the log). For the same paragraph: chunking 50 enabled = 52s, chunking disabled = 22s. I think this same text would have taken maybe 10-15s in very small chunks before I updated.
My AI assistant says: "Each chunk has a ~4-second initialization delay before processing begins, suggesting something expensive runs inside synthesize() that should only run once at startup (e.g., reference audio re-processing, CUDA context recreation, model reload)."
026-07-21 00:03:09 [INFO] server: Splitting text into chunks of size ~100.
2026-07-21 00:03:09 [INFO] utils: A single segment (length 179) exceeds chunk_size 100. It will form its own chunk.
2026-07-21 00:03:09 [INFO] utils: A single segment (length 168) exceeds chunk_size 100. It will form its own chunk.
2026-07-21 00:03:09 [INFO] utils: Text chunking complete. Generated 4 chunk(s).
2026-07-21 00:03:09 [INFO] server: Synthesizing chunk 1/4...
2026-07-21 00:03:09 [INFO] engine: Applying user-provided seed for generation: 9
2026-07-21 00:03:09 [INFO] engine: Global seed set to: 9
2026-07-21 00:03:13 [WARNING] chatterbox.tts_turbo: CFG, min_p and exaggeration are not supported by Turbo version and will be ignored.
27%|██▋ | 270/1000 [00:03<00:10, 71.84it/s]
S3 Token -> Mel Inference...
100%|██████████| 2/2 [00:00<00:00, 19.06it/s]
2026-07-21 00:03:25 [INFO] server: Synthesizing chunk 2/4...
2026-07-21 00:03:25 [INFO] engine: Applying user-provided seed for generation: 9
2026-07-21 00:03:25 [INFO] engine: Global seed set to: 9
2026-07-21 00:03:25 [WARNING] chatterbox.tts_turbo: CFG, min_p and exaggeration are not supported by Turbo version and will be ignored.
24%|██▎ | 236/1000 [00:03<00:10, 73.76it/s]
S3 Token -> Mel Inference...
100%|██████████| 2/2 [00:00<00:00, 21.47it/s]
2026-07-21 00:03:35 [INFO] server: Synthesizing chunk 3/4...
2026-07-21 00:03:35 [INFO] engine: Applying user-provided seed for generation: 9
2026-07-21 00:03:35 [INFO] engine: Global seed set to: 9
2026-07-21 00:03:35 [WARNING] chatterbox.tts_turbo: CFG, min_p and exaggeration are not supported by Turbo version and will be ignored.
13%|█▎ | 134/1000 [00:01<00:11, 73.68it/s]
S3 Token -> Mel Inference...
100%|██████████| 2/2 [00:00<00:00, 22.11it/s]
2026-07-21 00:03:43 [INFO] server: Synthesizing chunk 4/4...
2026-07-21 00:03:43 [INFO] engine: Applying user-provided seed for generation: 9
2026-07-21 00:03:43 [INFO] engine: Global seed set to: 9
2026-07-21 00:03:43 [WARNING] chatterbox.tts_turbo: CFG, min_p and exaggeration are not supported by Turbo version and will be ignored.
7%|▋ | 67/1000 [00:00<00:12, 73.28it/s]
S3 Token -> Mel Inference...
100%|██████████| 2/2 [00:00<00:00, 21.15it/s]
2026-07-21 00:03:51 [INFO] server: Smart stitching applied: 4 chunks, 20ms crossfades, 200ms pauses
2026-07-21 00:03:51 [INFO] utils: Encoded 1417004 bytes to 'wav' at 24000Hz in 0.023 seconds.
2026-07-21 00:03:51 [INFO] server: Successfully generated audio: tts_output_T07_E05_W05_20260721_000351.wav, 1417004 bytes, type audio/wav.
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.84 Driver Version: 595.84 CUDA Version: 13.2 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3090 On | 00000000:21:00.0 Off | N/A |
| 0% 33C P8 28W / 300W | 5259MiB / 24576MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
It was running great. The problem started after I upgraded my unRAID server where this is running via Docker. My guess is something downstream changed but I have been unable to figure it out in a few hours of troubleshooting. Maybe someone here has seen this before?
There are long pauses after each chunk (each progress bar in the log). For the same paragraph: chunking 50 enabled = 52s, chunking disabled = 22s. I think this same text would have taken maybe 10-15s in very small chunks before I updated.
My AI assistant says: "Each chunk has a ~4-second initialization delay before processing begins, suggesting something expensive runs inside synthesize() that should only run once at startup (e.g., reference audio re-processing, CUDA context recreation, model reload)."