| --- |
| language: |
| - ar |
| pretty_name: Arabic WebSocket TTS benchmark |
| --- |
| |
| # Arabic WebSocket TTS benchmark |
|
|
| A WebSocket load test for streaming TTS servers, and its results for |
| [voxcpm2-arabic-dialects-turbo](https://github.com/Speech-Next/voxcpm2-arabic-dialects-turbo). |
|
|
| | file | | |
| |---|---| |
| | `text.txt` | 3,592 Arabic sentences (mostly Egyptian), one per line; each request gets its own | |
| | `bench_voxcpm2_texts (1).py` | the benchmark: TTFB, total time, RTF, gap max, stalls per level of requests at once | |
| | `bench_voxcpm2_texts_underruns.py` | the same, plus underruns: a player starting at the first chunk runs dry > 50 ms | |
| | `results/` | runs of the scripts above, with a check against the server's own timestamps | |
|
|
| ```bash |
| python bench_voxcpm2_texts_underruns.py --levels 1,10,40,80,100,150 --dialect egyptian --texts-file text.txt |
| ``` |
|
|
| ## Results: voxcpm2-arabic-dialects-turbo, one A10 (2026-10-20) |
|
|
| `python -m nanovllm_voxcpm.entrypoints.ws_server` with its defaults; both scripts unchanged, three runs, the range |
| over the runs (`results/2026-10-20-a10`): |
|
|
| | requests at once | TTFA p99 | TTFB p95 | TTFB max | stalls | underruns | |
| |---:|---:|---:|---:|---:|---:| |
| | 1 | 24-29 ms | 178-185 ms | 178-185 ms | 0 | 0 | |
| | 10 | 29-36 ms | 187-195 ms | 187-196 ms | 0 | 0 | |
| | 40 | 58-70 ms | 219-239 ms | 219-245 ms | 0 | 0 | |
| | 80 | 69-73 ms | 241-256 ms | 241-256 ms | 0 | 0 | |
| | 100 | 76-81 ms | 257-270 ms | 257-270 ms | 0 | 0 | |
| | 150 | 94-148 ms | 275-330 ms | 288-330 ms | 0 | 0 | |
|
|
| - TTFB: the scripts' first chunk (4096 bytes, 128 ms of audio) at the client. The server holds the first chunk 150 ms |
| (`--pace-start-ms 150`, its default): the time the engine needs for no stalls. |
| - TTFA: the engine's first audio after the text reached the server, from the server's timestamps. |
| - Check (`verify.py`, `verify.md`): the stalls recomputed from the server's send times equal the scripts' (0 / 0 at |
| every level), and the client's TTFB percentiles are the server's plus 0-22 ms (the WebSocket hops). |
| - 150 at once is the limit of one A10: at 200 at once, 24-84 stalls per run (no underruns). |
| - No audio here: the generated speech clones a real speaker's voice. |
|
|
