Choose theme
AI Resources
VoxCPM2
Generate narration from text, describe a new voice, or use a permitted recording as the reference for speech generation.
VoxCPM2 is OpenBMB's speech model, with documented support for 30 languages and 48 kHz output. Its Python library and demo expose different ways to choose the voice and delivery. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
A multilingual TTS model
OpenBMB describes a tokenizer-free model that turns written text into speech across 30 languages, with 48 kHz output.
Why it stands out
Voice design and cloning controls
Describe a voice without a recording, clone from reference audio without a transcript, or supply prompt audio with its exact transcript for continuation-style generation.
Availability
Model, docs, and demos
Hugging Face hosts the checkpoint; the repository, usage guide and demo provide generation examples and setup instructions.
Why it matters
What makes it useful
Turn a written passage into a WAV file, or play generated audio chunks in order through generate_streaming(). The usage guide describes splitting incoming text into sentences for streaming output; this does not support text arriving token by token while speech is generated.
What to know
Where it fits
Without reference audio, Voice Design uses a description in parentheses before the spoken text. For cloning, reference_wav_path supplies the voice without a transcript. The Hi-Fi example also supplies prompt_wav_path and the recording's exact prompt_text. Choose the input route around the reference material you can legitimately provide.
Notable points
What stands out
Controllable cloning separates the reference voice from delivery instructions: a parenthesized prompt can request a different pace or emotion. The guide says the demo's Hi-Fi mode ignores that control instruction, so the transcript-assisted route and the style-controlled route are different choices.
Before using
What to review
Follow the selected release and device instructions. Release 2.0.3 specifies Python 3.10 to below 3.13 and adds CPU, MPS and CUDA device selection; the model card reports approximately 8 GB of memory for generation, not a guarantee for every setup.
The usage guide suggests clean reference audio of 5–30 seconds and recommends splitting long passages into shorter segments to reduce instability. The model card notes that output can vary between runs and performance varies across languages.
Write audio using model.tts_model.sample_rate. Older VoxCPM examples use different output rates; the VoxCPM2 checkpoint outputs 48 kHz.
Use reference voices you have permission to use. The model card prohibits impersonation, fraud and disinformation and asks for generated speech to be labelled as AI-generated; a cloning option does not establish voice rights.
Reader fit
Who may find it relevant
Creators preparing multilingual narration and experimenting with described voices or permitted reference recordings.
Developers connecting speech generation or ordered audio chunks to an application.
This is a model and its tooling; a finished dialogue editor or general assistant requires another application layer.
Editorial note
Why LifeHubber lists it
VoxCPM2 is worth considering for creators preparing speech in a supported dialect. Its cookbook connects that generation choice to dialect-specific script writing, beyond selecting a voice or delivery style. For Cantonese or another supported dialect, the cookbook recommends preparing the spoken text in that dialect's own vocabulary and keeping the dialect instruction simple. Its Cantonese example contrasts that wording with a standard-Mandarin version: a Cantonese control label alone is not the script preparation it describes. This gives a writer a concrete text-editing step beyond choosing a speaker or cloning mode. This does not guarantee pronunciation or consistent quality across dialects.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
What to explore next
Move from generation to an application endpoint.
If your next task is serving repeated speech requests, OpenBMB documents a VoxCPM2 integration with vLLM-Omni. Continue with the serving framework, then use the linked model-specific recipe to check your platform and endpoint.
More in Speech Models
Keep browsing this category
Explore more speech model resources.
Fish Audio S2 Pro
fishaudio/s2-pro
A text-to-speech model with in-script control over pauses, emphasis and mood.
AuK
Tencent-Hunyuan/AuK
A 1.5B speech model for instruction-guided text-to-speech, content and acoustic editing, paralinguistic changes, speech enhancement, and source separation, with public code, weights, demos, ComfyUI nodes, and fine-tuning materials.
TADA
HumeAI/tada
A speech-language model that aligns speech and text into a single synchronized stream.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
