VoxCPM
VoxCPM2 is a tokenizer-free open-source TTS model for 30-language speech synthesis, Voice Design, and 48kHz voice cloning on your own GPU.
For teams already running GPUs, VoxCPM2 is the open-weight TTS release worth benchmarking first: three distinct voice-creation modes, 30 languages, 48kHz output, and an Apache-2.0 license that kills the per-character bill. The catch is operational — Python 3.10+, PyTorch 2.5+, CUDA 12+, and an RTX 4090-class card if you want that ~0.3 RTF streaming figure. If you need a hosted endpoint by Friday, a managed API still wins on time-to-first-audio.
Verified 3d ago · liveness 76/100 · cite: rightaichoice.com/tools/voxcpm
- Developers building multilingual voice assistants without per-character API costs
- Game and media studios designing dynamic character voices from text descriptions
- Researchers exploring tokenizer-free speech synthesis and custom model fine-tuning
- Teams needing commercially usable TTS with full control over weights and data path
- Non-technical users wanting a turnkey hosted TTS API with no setup
- Teams without CUDA-capable GPUs who need real-time streaming out of the box
- Projects that can't budget engineering time for Python, CUDA, and model serving
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip VoxCPM if you have no CUDA GPU and no engineering bandwidth for Python environments and weight downloads, and need a hosted TTS endpoint you can call today.
The license is free but the GPU isn't — real-time streaming at ~0.3 RTF assumes an RTX 4090-class card, and cloud GPU rental adds a recurring bill that a per-character API would have folded into usage.
VoxCPM costs $0 — Apache-2.0 weights and code, no per-character fee, no seat minimum. That makes it cheaper than any metered hosted TTS at volume, including ElevenLabs and OpenAI TTS, once you're past the point where per-character billing exceeds a GPU's monthly cost. Below that volume, or without a GPU already on hand, hosted APIs are the cheaper total spend because they price in the ops work you'd otherwise do yourself. Team size barely matters; what matters is whether you already run CUDA
In short
VoxCPM — VoxCPM2 is a tokenizer-free open-source TTS model for 30-language speech synthesis, Voice Design, and 48kHz voice cloning on your own GPU. Best for Developers building multilingual voice assistants without per-character API costs, Game and media studios designing dynamic character voices from text descriptions, Researchers exploring tokenizer-free speech synthesis and custom model fine-tuning. Free to use.
What's new in VoxCPM
Checked 2 days agoAcross the latest 4 updates: 1 feature update, 2 launches and 1 news mention.
VoxCPM2: 2B model with 30 languages, Voice Design, Controllable & Ultimate Cloning, 48kHz output
Major release built on a MiniCPM-4 backbone with AudioVAE V2 and vLLM-Omni acceleration, shipping 2B-parameter Apache-2.0 weights covering 30 languages and three voice-creation paths.
Open-source VoxCPM1.5 weights with SFT & LoRA fine-tuning
VoxCPM1.5 weights were released alongside SFT and LoRA fine-tuning support, and the project reached #1 on GitHub Trending.
VoxCPM Technical Report release
A technical report was published documenting the tokenizer-free diffusion autoregressive architecture and its training approach.
Open-source VoxCPM-0.5B weights
The smaller 0.5B-parameter VoxCPM model was released and reached #1 on HuggingFace Trending.
What people actually say about VoxCPM — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
15 mentions across 4 sources (Hacker News, Product Hunt, GitHub, Lemmy) · researched Jul 3, 2026.
Average across the 4 sources that answered — each source counts once, not each post.
- +Tokenizer-free architecture produces more natural prosody and intonation than discrete-token TTS systems.
- +Voice cloning from a short audio sample captures breathing, accents, and emotion details very accurately.
- +Supports 30 languages and outputs at 48kHz with real-time streaming inference potential.
- +Completely free and open-source under Apache-2.0 with no vendor lock-in.
- +Active community on GitHub with 32k+ stars and frequent updates.
- −CLI documentation is inconsistent — --device cpu flag doesn't work despite being shown in tutorials.
- −Polish and some non-English languages suffer from abrupt voice cutoffs at word ends.
- −CUDA errors on certain NVIDIA GPUs (e.g., L20) disrupt inference reliability.
- −Artifact from reference audio bleeds into output during transcript-guided cloning.
- −Fine-tuning new languages lacks clear documentation — only Chinese and English well-supported.
- • Requires powerful GPU for acceptable inference speed—costs may be incurred for cloud GPU rentals
- • No official hosting—users must provide their own infrastructure
Viability Score
How well maintained and how widely used is VoxCPM? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Tokenizer-free diffusion autoregressive TTS architecture (no discrete tokenization)
- VoxCPM2, 2B parameters on a MiniCPM-4 backbone, trained on 2M+ hours of speech
- 30-language synthesis with no language tags required
- Chinese dialect support including 四川话, 粤语, 吴语, 东北话, 河南话, 陕西话, 闽南话
- Voice Design from a natural-language voice description, no reference audio
- Controllable Cloning with emotion, pace, and expression guidance
- Ultimate Cloning from reference audio plus its transcript
- 48kHz studio-quality output via AudioVAE V2
- Built-in super-resolution accepting 16kHz reference audio
- Context-aware prosody inferred from text content
- Real-time streaming at RTF ~0.3 on an NVIDIA RTX 4090
- vLLM-Omni acceleration at RTF ~0.13 with PagedAttention
- OpenAI-compatible API served through vLLM-Omni
- On-device inference via llama.cpp-omni
- SFT and LoRA fine-tuning, with a LoRA fine-tuning web UI
About VoxCPM
VoxCPM is an open-source text-to-speech system from OpenBMB that generates continuous speech representations through an end-to-end diffusion autoregressive architecture, skipping discrete tokenization entirely. VoxCPM2, released April 2026, is a 2B-parameter model on a MiniCPM-4 backbone trained on over 2 million hours of multilingual speech data, covering 30 languages with no language tags required, plus Chinese dialects including 四川话, 粤语, 吴语, 东北话, and 闽南话. The repo sits at roughly 39.3k GitHub stars under Apache-2.0. It ships three voice-creation paths. Voice Design builds a brand-new voice from a natural-language description of gender, age, tone, emotion, and pace with no reference audio. Controllable Cloning copies a timbre from a short clip while letting you steer emotion, pace, and expression. Ultimate Cloning takes reference audio plus its transcript to preserve every vocal detail. Context-aware synthesis infers prosody from the text itself. Output lands at 48kHz through AudioVAE V2, whose asymmetric encode/decode design and built-in super-resolution accept 16kHz reference clips without an external upsampler. It ships as a Python library (pip install voxcpm), a CLI, a bundled web demo, and a LoRA fine-tuning web UI. Streaming is realistic for production: RTF as low as ~0.3 on an NVIDIA RTX 4090, dropping to ~0.13 with Nano-vLLM or vLLM-Omni, the latter serving an OpenAI-compatible API with PagedAttention. On-device inference runs through llama.cpp-omni. Weights and code are Apache-2.0, so commercial deployment carries no per-character fee — you supply the GPU and the engineering time and get full control of the model and data path in return. Against hosted APIs like ElevenLabs or OpenAI TTS, VoxCPM trades time-to-first-audio for ownership; against other open-weight TTS, its range of cloning modes and language coverage is the differentiator.
Behind the Verdict
VoxCPM2 is what open-weight TTS looks like when a lab commits to a real release cadence. Since 0.5B in September 2025 and VoxCPM1.5 in December, the April 2026 jump to a 2B MiniCPM-4 backbone with 30 languages and 48kHz output is the version most teams should target — the earlier weights are now legacy. Pick this when you have GPU capacity and an engineer who can own the serving stack. The vLLM-Omni path is the one that matters in production: PagedAttention plus an OpenAI-compatible API means a lot of existing client code works unchanged, and RTF near 0.13 leaves headroom for concurrent streams on a single card. ModelScope download support helps if HuggingFace is slow from your region. Where it bites: setup is real work. You are managing Python ≥3.10 (<3.13), PyTorch ≥2.5.0, CUDA ≥12.0, model weights, and a denoiser toggle before you hear a single word. Non-technical teams will bounce off this in the first hour. There is also no escaping the hardware line — no CUDA GPU means no real-time streaming, full stop. Compared to a hosted API such as ElevenLabs or OpenAI TTS, you are trading convenience for control. Those services get you audio today and bill per character forever. VoxCPM gets you a fixed GPU cost and complete say over the data path, which matters if your scripts or reference voices can't leave your infrastructure. Against other open-weight TTS, the differentiator is the three-mode split. Voice Design alone is worth the install if you are prototyping characters — you describe a voice and get audio back with no reference clip to source or clear. Controllable Cloning is the practical middle: copy a timbre from a short sample, then steer emotion and pace per line. Ultimate Cloning with transcript is the fidelity option when a brand voice must be reproduced
Researching VoxCPM? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas VoxCPM actually fits — and what changes day-one when you adopt it.
Install voxcpm via pip, open the bundled web demo, and write voice descriptions for each character — 'older male, gravelly, slow and weary' — to generate lines with Voice Design and no reference audio.
Outcome: A full cast of distinct voices with no per-line API cost, generated locally and iterated on until each character sounds right.
Feed the same source script through VoxCPM2 in each of the 30 supported languages without tagging the language, then use Controllable Cloning on one approved voice to steer pace and emotion per locale.
Outcome: Consistent voice-over across every target market from a single self-hosted engine, with no per-character bill scaling with the language count.
Stand up vLLM-Omni on an RTX 4090, serve VoxCPM2 behind its OpenAI-compatible API with PagedAttention, and point existing TTS client code at the new base URL.
Outcome: The ~0.13 RTF self-hosted endpoint drops in without rewriting application code, and per-character spend stops.
Use Cases
- Generate natural-sounding multilingual speech for audiobook narration across 30 languages.
- Clone a voice from a short reference clip for personalized voice assistants.
- Create unique voice characters for video games from text descriptions alone.
- Stream real-time speech output for interactive voice applications using vLLM-Omni.
- Fine-tune the 2B model on domain-specific voice data with SFT or LoRA.
- Run offline TTS on your own hardware for data-sensitive workflows.
- Produce voice-over localization for 30 languages without per-character API fees.
- Swap a hosted TTS endpoint for an OpenAI-compatible self-hosted server.
Models Under the Hood
as of 2026-10-17
Limitations
- VoxCPM2 needs a CUDA-compatible GPU for real-time streaming — RTF is ~0.3 on an RTX 4090, so consumer or CPU-only hardware pushes latency up.
- Fine-tuning and accelerated serving via vLLM-Omni or Nano-vLLM involve real setup: Python environments, model weight downloads, and Docker.
- There is no managed cloud service; you self-host and you own the uptime.
- Voice Design and Ultimate Cloning reward careful prompt engineering and good reference-audio selection, and results vary with both.
- The Apache-2.0 license covers the weights and code, but it does not grant you rights to the voices you clone — that clearance is entirely on you.
as of 2026-10-08
Verification history
We have re-verified VoxCPM 10 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 10 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published VoxCPM tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Open Source (Apache-2.0)
$0
Ideal for
Teams with a CUDA GPU and an engineer who can run a serving stack, from solo developers to studios shipping commercial TTS at volume.
What this tier adds
Single free tier — Apache-2.0 weights and code with no usage cap, no per-character fee, and commercial use permitted.
Where the pricing makes sense
The company stage and team size where VoxCPM's pricing actually pencils out — and where peers do it cheaper.
VoxCPM costs $0 — Apache-2.0 weights and code, no per-character fee, no seat minimum. That makes it cheaper than any metered hosted TTS at volume, including ElevenLabs and OpenAI TTS, once you're past the point where per-character billing exceeds a GPU's monthly cost. Below that volume, or without a GPU already on hand, hosted APIs are the cheaper total spend because they price in the ops work you'd otherwise do yourself. Team size barely matters; what matters is whether you already run CUDA
Setup time & first value
How long it actually takes to get something useful out of VoxCPM — broken out by persona, not the marketing-page minute.
Developers already comfortable with Python and CUDA: roughly an hour to pip install voxcpm, pull the weights, and generate first audio in the web demo. Platform engineers going to production: half a day to a day to stand up vLLM-Omni or Nano-vLLM with Docker and get the OpenAI-compatible endpoint serving. Fine-tuning with SFT or LoRA is a separate multi-day effort to prepare a dataset and run
Switching to or from VoxCPM
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From ElevenLabs: point existing client calls at the vLLM-Omni OpenAI-compatible endpoint and map each hosted voice to a VoxCPM2 cloning or Voice Design preset.
- →From OpenAI TTS: keep the OpenAI SDK and change only the base URL, then validate prosody on your own text corpus before cutting over.
- →From VoxCPM1.5: Ultimate Cloning works the same way, so existing reference-audio-plus-transcript pipelines carry over and you gain 48kHz output, Voice Design, and 30-language coverage.
- →From a commercial TTS vendor: export your audio assets, recreate each voice via Ultimate Cloning from a licensed reference clip, and retire the per-character contract.
- ↗To ElevenLabs: re-create each Voice Design or cloned voice in the hosted library and accept per-character pricing in exchange for zero serving ops.
- ↗To OpenAI TTS: map scripts to the hosted voice set and drop the GPU and serving stack entirely, trading voice-design flexibility for a managed endpoint.
- ↗To VoxCPM1.5: revert the clone pipeline if you depend on the older 0.5B-class footprint and can forgo 48kHz output and Voice Design.
Integrations
Resources & Guides
Tutorials & Learning
YouTube returned 6 videos for “VoxCPM”, and we withheld 6: 6 could not be judged, because “VoxCPM” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about VoxCPM.
Official links
Tools that pair well with VoxCPM
Common stack mates teams adopt alongside VoxCPM, with the specific reason each pairing earns its keep.
ComfyUI VoxCPM
ComfyUI VoxCPM brings open-source multilingual text-to-speech, voice cloning and voice design onto your own GPU.
OmniVoice Studio
OmniVoice Studio is an open-source desktop studio for local voice cloning, voice design, dubbing, dictation and audiobook production.
Inworld
Inworld delivers realtime text-to-speech, speech-to-text, and open-model inference on one API for consumer voice apps.
Featured Head-to-Head Comparisons
Voxcpm vs Soniox
If you need a production-ready, compliant, multilingual STT+TTS+translation API with sub-200ms latency, choose Soniox. If you want a free, open-source TTS model with voice cloning and plan to self-host on a GPU, choose VoxCPM. Soniox wins for enterprise-grade voice applications; VoxCPM wins for budget-friendly TTS experimentation.
Voxcpm vs Retell Ai
Choose Retell AI if you need a turnkey, cloud-based voice agent platform for phone automation at scale with low latency and rich integrations, even at a cost. Choose VoxCPM if you prioritize zero licensing fees, full control over deployment, multilingual TTS, and voice cloning, and you have the technical capacity to self-host on GPU infrastructure. They serve fundamentally different needs: Retell is an all-in-one call center solution, while VoxCPM is a highly capable open-source speech synthesis model.
Voxcpm vs Voiceitt
Choose Voiceitt if you need speech recognition for non-standard speech patterns (e.g., cerebral palsy, heavy accents) and value integrations with Webex, Teams, and Alexa. Choose VoxCPM if you want open-source TTS with voice cloning and multilingual support, and have the technical ability to run a 2B model locally on a GPU. They serve fundamentally different needs.
Alternatives to VoxCPM
View all
ComfyUI VoxCPM
ComfyUI VoxCPM brings open-source multilingual text-to-speech, voice cloning and voice design onto your own GPU.
OmniVoice Studio
OmniVoice Studio is an open-source desktop studio for local voice cloning, voice design, dubbing, dictation and audiobook production.
Frequently Asked Questions
Categories
Best-of guides
Topics
Used VoxCPM? Help shape our editorial sentiment research.