Text-to-Speech
VoxCPM
Safetensors
English
tts
lora
single-speaker
audio
File size: 6,987 Bytes
e1636a7
 
 
 
 
 
 
 
8283dfb
 
e1636a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d959d49
 
 
e1636a7
 
 
 
 
 
 
 
 
 
 
 
 
 
8283dfb
e1636a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d959d49
 
e1636a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
---
license: apache-2.0
language:
- en
library_name: voxcpm
pipeline_tag: text-to-speech
base_model: openbmb/VoxCPM2
base_model_relation: finetune
datasets:
- alejandro-hp/AlephTTS-data
tags:
- text-to-speech
- tts
- voxcpm
- lora
- single-speaker
- audio
---

# AlephTTS

**AlephTTS** is a text-to-speech model that speaks in the voice of **Aleph One-Half**, a Twitch streamer.
It is a fine-tune of [VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) (2B parameters,
48 kHz output) on ~47 minutes of the voice owner's own recordings. The voice is built into the weights:
no reference audio is needed.

## Samples

All samples were generated by this model (not recorded).

| | |
|---|---|
| "Hey everyone, Aleph One-Half here, and welcome back to the stream!" | <audio controls src="https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/samples/01_greeting.wav"></audio> |
| "No way! That's the rarest item in the entire game!" | <audio controls src="https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/samples/02_hype.wav"></audio> |
| "The derivative of x cubed is three x squared." | <audio controls src="https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/samples/03_math.wav"></audio> |
| "Thanks so much for the follow, seriously, it means a lot." | <audio controls src="https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/samples/04_chat.wav"></audio> |
| "That's it for tonight, everyone. Get some rest, and I'll see you next time." | <audio controls src="https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/samples/05_goodbye.wav"></audio> |

## Usage

```bash
pip install voxcpm soundfile   # Python 3.10–3.12, PyTorch >= 2.5, NVIDIA GPU with ~6 GB free VRAM
```

```python
from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained("alejandro-hp/AlephTTS", load_denoiser=False)

wav = model.generate(text="Hey chat, welcome back to the stream!",
                     cfg_value=2.0, inference_timesteps=10)
sf.write("out.wav", wav, model.tts_model.sample_rate)   # 48 kHz mono

# Streaming (first audio in well under a second):
for chunk in model.generate_streaming(text="Okay chat, here we go.",
                                      cfg_value=2.0, inference_timesteps=10):
    ...  # float32 numpy chunks at model.tts_model.sample_rate
```

Tips:
- Don't pass `reference_wav_path` / `prompt_wav_path`; the voice is already built in.
- Write numbers and math out in words ("x squared plus y squared equals one", "one in four"),
  or pass `normalize=True`.
- Punctuation shapes delivery: commas and periods add natural pauses, "!" adds energy.
- For long text, generate sentence by sentence.
- Load the model once and keep it in memory; the first `generate` call after loading includes a warm-up.
- **Low on VRAM?** Load with `VoxCPM.from_pretrained("alejandro-hp/AlephTTS", load_denoiser=False, optimize=False)`.
  This skips `torch.compile`/CUDA graphs: peak VRAM drops to ~5.6 GB at the same speed on an RTX 3070 Ti,
  which helps on 8 GB cards shared with a browser, game or OBS.
- `seed=N` (reproducible output) requires a voxcpm release newer than 2.0.3.

## Model details

| | |
|---|---|
| Base model | [openbmb/VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) (Apache-2.0) |
| Method | LoRA (rank 32, alpha 32) on the LM and diffusion transformer attention projections, merged into the base weights |
| Language | English (American) |
| Speaker | One: Aleph One-Half (the uploader) |
| Output | 48 kHz mono |
| Weights | bf16 safetensors, ~4.6 GB, plus the unchanged base audio codec (`audiovae.pth`) |

### Training data
[alejandro-hp/AlephTTS-data](https://huggingface.co/datasets/alejandro-hp/AlephTTS-data) (CC0-1.0):
~47 minutes (854 clips) of purpose-made recordings by the voice owner, captured with a Shure MV7+ at
48 kHz / 24-bit: phonetically balanced read sentences (CMU ARCTIC prompts), conversational and
stream-style lines (greetings, reactions, hype, calm), math explanations, and lines containing the
username. Takes were checked with automatic transcription, trimmed, and screened for mid-phrase
hesitations to favor fluent delivery.

### Training procedure
LoRA fine-tuning with VoxCPM's official training code (adapted to fit an 8 GB GPU: frozen base weights in
bf16, gradient checkpointing), learning rate 1e-4, effective batch 16, 3000 steps (~57 epochs). The
checkpoint was selected by blind listening tests.

### Evaluation
- Blind listening test by the voice owner (1–5 scale, 10 sentences): **5.0**, tied with real recordings.
- Word error rate (Whisper large-v3) on 24 test sentences, including unseen sentences: 0%.
- Speaker similarity (ECAPA-TDNN) to held-out real recordings: 0.73–0.76 (real-vs-real: 0.77).
- On an RTX 3070 Ti: ~5.8 GB VRAM (~5.6 GB with `optimize=False`), real-time factor ~0.74,
  first streaming chunk in under 0.1 s.

## Intended use and limitations

**Intended use:** speech in the voice of Aleph One-Half for the voice owner's own content (for example
stream alerts and narration), and for research or experimentation where outputs are clearly labeled as
synthetic.

**Out of scope / not allowed:** impersonating Aleph One-Half or anyone else, deceiving listeners about
who is speaking, fraud, harassment, or any use that violates the
[Hugging Face content policy](https://huggingface.co/content-policy) or applicable law. Please label
AI-generated audio as such.

**Consent:** the voice in this model belongs to the uploader, who recorded the training data
specifically for this model and consents to its publication under the terms above.

**Limitations:** English only; trained on under an hour of a single speaker, so very unusual words, names
or symbols may be mispronounced; extreme emotions beyond the recorded styles may sound less natural.
Inherits the limitations of VoxCPM2.

## License and attribution

Apache-2.0 (see `LICENSE`). This is a derivative of **VoxCPM2 by OpenBMB / the VoxCPM Team**, licensed
under Apache-2.0.

Modifications relative to VoxCPM2: `model.safetensors` contains the base weights with a LoRA fine-tune
merged in (2026-11-16). `config.json`, the tokenizer files and `audiovae.pth` are unchanged copies from
the base model. `export_info.json` documents the fine-tune configuration and file checksums.

## Citation

If you use this model, please cite VoxCPM:

```bibtex
@article{voxcpm2_2026,
  title   = {VoxCPM2: Tokenizer-Free TTS for Multilingual Speech Generation, Creative Voice Design, and True-to-Life Cloning},
  author  = {VoxCPM Team},
  journal = {GitHub},
  year    = {2026},
}

@article{voxcpm2025,
  title   = {VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning},
  author  = {Zhou, Yixuan and Zeng, Guoyang and Liu, Xin and Li, Xiang and
             Yu, Renjie and Wang, Ziyang and Ye, Runchuan and Sun, Weiyue and
             Gui, Jiancheng and Li, Kehan and Wu, Zhiyong and Liu, Zhiyuan},
  journal = {arXiv preprint arXiv:2509.24650},
  year    = {2025},
}
```