Text-to-Speech
VoxCPM
Safetensors
English
tts
lora
single-speaker
audio
alejandro-hp commited on
Commit
d959d49
·
verified ·
1 Parent(s): e1636a7

Model card: add low-VRAM tip (optimize=False)

Browse files
Files changed (1) hide show
  1. README.md +5 -1
README.md CHANGED
@@ -63,6 +63,9 @@ Tips:
63
  - Punctuation shapes delivery: commas and periods add natural pauses, "!" adds energy.
64
  - For long text, generate sentence by sentence.
65
  - Load the model once and keep it in memory; the first `generate` call after loading includes a warm-up.
 
 
 
66
  - `seed=N` (reproducible output) requires a voxcpm release newer than 2.0.3.
67
 
68
  ## Model details
@@ -92,7 +95,8 @@ checkpoint was selected by blind listening tests.
92
  - Blind listening test by the voice owner (1–5 scale, 10 sentences): **5.0**, tied with real recordings.
93
  - Word error rate (Whisper large-v3) on 24 test sentences, including unseen sentences: 0%.
94
  - Speaker similarity (ECAPA-TDNN) to held-out real recordings: 0.73–0.76 (real-vs-real: 0.77).
95
- - On an RTX 3070 Ti: ~5.8 GB VRAM, real-time factor ~0.74, first streaming chunk in under 0.1 s.
 
96
 
97
  ## Intended use and limitations
98
 
 
63
  - Punctuation shapes delivery: commas and periods add natural pauses, "!" adds energy.
64
  - For long text, generate sentence by sentence.
65
  - Load the model once and keep it in memory; the first `generate` call after loading includes a warm-up.
66
+ - **Low on VRAM?** Load with `VoxCPM.from_pretrained("alejandro-hp/AlephTTS", load_denoiser=False, optimize=False)`.
67
+ This skips `torch.compile`/CUDA graphs: peak VRAM drops to ~5.6 GB at the same speed on an RTX 3070 Ti,
68
+ which helps on 8 GB cards shared with a browser, game or OBS.
69
  - `seed=N` (reproducible output) requires a voxcpm release newer than 2.0.3.
70
 
71
  ## Model details
 
95
  - Blind listening test by the voice owner (1–5 scale, 10 sentences): **5.0**, tied with real recordings.
96
  - Word error rate (Whisper large-v3) on 24 test sentences, including unseen sentences: 0%.
97
  - Speaker similarity (ECAPA-TDNN) to held-out real recordings: 0.73–0.76 (real-vs-real: 0.77).
98
+ - On an RTX 3070 Ti: ~5.8 GB VRAM (~5.6 GB with `optimize=False`), real-time factor ~0.74,
99
+ first streaming chunk in under 0.1 s.
100
 
101
  ## Intended use and limitations
102