Instructions to use alejandro-hp/AlephTTS with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- VoxCPM
How to use alejandro-hp/AlephTTS with VoxCPM:
import soundfile as sf from voxcpm import VoxCPM model = VoxCPM.from_pretrained("alejandro-hp/AlephTTS") wav = model.generate( text="VoxCPM is an innovative end-to-end TTS model from ModelBest, designed to generate highly expressive speech.", prompt_wav_path=None, # optional: path to a prompt speech for voice cloning prompt_text=None, # optional: reference text cfg_value=2.0, # LM guidance on LocDiT, higher for better adherence to the prompt, but maybe worse inference_timesteps=10, # LocDiT inference timesteps, higher for better result, lower for fast speed normalize=True, # enable external TN tool denoise=True, # enable external Denoise tool retry_badcase=True, # enable retrying mode for some bad cases (unstoppable) retry_badcase_max_times=3, # maximum retrying times retry_badcase_ratio_threshold=6.0, # maximum length restriction for bad case detection (simple but effective), it could be adjusted for slow pace speech ) sf.write("output.wav", wav, 16000) print("saved: output.wav") - Notebooks
- Google Colab
- Kaggle
|
Download README.md from alejandro-hp/AlephTTS: direct link, hf CLI and curl.
- Browser
- Download file 6.99 kB
-
https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/README.md
- Command line
-
hf download hf://alejandro-hp/AlephTTS/README.md
-
curl -L -o README.md https://huggingface.co/alejandro-hp/AlephTTS/resolve/main/README.md
6.99 kB
| 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}, | |
| } | |
| ``` | |