Nighthawk/Qwen3-TTS-streaming
1
0
Fork 0
mirror of https://github.com/Nighthawk42/Qwen3-TTS-streaming.git synced 2026-08-30 08:52:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #7 from rekuenkdr/fix/repetition-penalty-and-finetune-fixes

Browse Source 
docs: add repetition_penalty param to README with streaming context
This commit is contained in:
rekuenkdr
2026-02-07 10:42:41 +01:00
committed by GitHub
parent cecfe01b67 f88566ac7e
commit 1275736834
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+5 -3
View File
@@ -10,8 +10,10 @@ From [dffdeeq/Qwen3-TTS-streaming](https://github.com/dffdeeq/Qwen3-TTS-streamin
- `torch.compile` + CUDA graphs optimization
Added in this fork:
- **Two-phase streaming** - faster first-chunk latency (2.75x improvement)
- **Audio quality fixes** - click-free chunk blending with Hann crossfade
- **Two-phase streaming** - faster first-chunk latency
- **Repetition penalty for streaming** - prevents token loops that cause looping audio and runaway generation. Defaults to 1.0 (disabled) because streaming generates frame-by-frame with CUDA graph constraints where repetition manifests differently than the non-streaming path (which defaults to 1.05)
- **Multiple EOS token detection** - broader termination coverage for reliable generation stopping
- **Hann window crossfade** - click-free chunk boundaries with proper fade-in/fade-out
## Two-Phase Streaming
@@ -165,7 +167,7 @@ for chunk, sr in model.stream_generate_voice_clone(
| `first_chunk_emit_every` | 0 | Phase 1 emit interval (0 = disabled) |
| `first_chunk_decode_window` | 48 | Phase 1 decode window |
| `first_chunk_frames` | 48 | Switch to phase 2 after N frames |
| `use_optimized_decode` | True | Use torch.compile/CUDA graph optimized decode |
| `repetition_penalty` | 1.0 | Penalizes repeated tokens (1.0 = disabled). Increase toward 1.05 if you hear looping artifacts |
## Installation