Phonon vs VoxCPM2: What 64× More Parameters Actually Buys in TTS

VoxCPM2 is a 2 billion parameter speech model from OpenBMB, released in April 2026, that speaks 30 languages at 48 kHz and can invent a voice from a written description. Phonon 31M is a 31 million parameter model from Gradium, released on 20 October 2026, that speaks English at 24 kHz in 37 fixed voices and fits in 64.5 MB. The parameter ratio is about 64 to 1, and the surprising part is not that the bigger model does more. It is that on the one number both vendors publish — English word error rate on Seed-TTS — the smaller model claims the better score.

That claim is almost certainly not a fair comparison, and understanding why tells you more about the state of open TTS than either model's launch post does. VoxCPM2 and Phonon 31M are both Apache 2.0, both downloadable, and both aimed at teams that want speech without a per-character bill. They are aimed at completely different hardware.

Two different jobs wearing the same label

VoxCPM2 is a general-purpose multilingual speech generator. It is tokenizer-free and diffusion-autoregressive, operating in the AudioVAE V2 latent space through four stages the authors label LocEnc, TSLM, RALM and LocDiT, on a MiniCPM-4 backbone. It supports voice design from a natural-language description, controllable cloning from a short reference, and continuation-style cloning that uses both a reference clip and its transcript. It outputs 48 kHz audio and accepts 16 kHz reference audio.

Phonon 31M is a single-language, single-purpose synthesizer. It is a flow-matching transformer over Mimi latents, 80 millisecond frames at a time, decoded to 24 kHz mono. It has no voice design, no cloning from a reference clip in the documented API, and no language detection. It has a voice catalogue, a text input, and a runtime that starts in single-digit milliseconds.

Calling both "open TTS" is accurate and unhelpful. One is a speech research platform; the other is an embedded component.

Where the 64× goes

The parameter budget buys four things, and all four are visible in the spec sheet rather than the listening test.

• Languages — 30 plus nine Chinese dialects for VoxCPM2; English only for Phonon 31M.

• Output rate — 48 kHz native for VoxCPM2; 24 kHz mono for Phonon 31M.

• Voice creation — description-based voice design and three cloning modes for VoxCPM2; 37 built-in voices plus a custom-voice catalogue for Phonon 31M.

• Hardware floor — about 8 GB of VRAM on an RTX 4090 for VoxCPM2; 64.5 MB and a CPU core for Phonon 31M.

• Speed — roughly 0.30 real-time factor on an RTX 4090 for VoxCPM2, or 0.13 with Nano-vLLM; 26 to 41 times faster than real time on a laptop core for Phonon 31M, according to Gradium's README.

• Licence — Apache 2.0 for both.

The WER comparison is not a comparison

VoxCPM2's technical report lists an English WER of 1.84 on Seed-TTS-eval, with a speaker similarity of 75.3. Phonon's model card lists 0.51 on Seed-TTS for Phonon 31M. Read side by side, that looks like a rout: a 31M model beating a 2B model on intelligibility by a factor of three.

It is not. The two numbers come from different harnesses, likely different reference sets, and almost certainly different transcription models used to score the output. Gradium's published methodology for its earlier evaluation used the English portion of Seed-TTS — 1,008 utterances — transcribed with whisper-large-v3 and scored with edit distance. OpenBMB's technical report does not use the identical pipeline in the figures surfaced on its model page. Neither vendor has run the other's checkpoint through its own harness, and no third party has published a head-to-head.

What the numbers do support is a weaker, still useful claim: both models are in the low-single-digit WER range on standard English test sets, which is to say both are intelligible enough for production read-aloud work. The idea that 31M parameters cannot produce clean English is simply false in 2026. The idea that 31M parameters can match a 2B model across 30 languages is also false, and neither vendor claims it.

There is a second reason to distrust the headline: Gradium's own documents disagree with each other. The release blog gives Phonon 31M a hard-set WER of 1.70%, the model card gives 1.5%, and the runtime README publishes no WER at all. When a vendor's three documents differ by 13% on the same metric, the right response is to treat the figure as a range.

The hardware floor is the real decision

VoxCPM2 wants a GPU. The reference configuration is about 8 GB of VRAM, and the model runs at roughly a third of real time on an RTX 4090 without an optimised serving stack. On an Apple M4 Pro with Q8_0 quantization the real-time factor is about 1.76, meaning it takes 1.76 seconds of compute to produce one second of audio — slower than real time, and not viable for live conversation on that hardware. With Nano-VLLM the factor drops to about 0.13, which is fast, but that is a serving stack with its own operational cost.

Phonon 31M's floor is a CPU core. Gradium's README lists 26 times real time on one MacBook Air M5 core, 22 times on an iPhone 16 Pro Neural Engine, and 2.3 times on a Raspberry Pi 5. The release blog's figures are different — 11.6 times on one core, 3.7 times on the Pi — but every version of the number describes a device that fits in a hand.

This is not a quality argument. A 48 kHz VoxCPM2 render through a GPU will sound better than a 24 kHz Phonon render, and it should. The question is whether your product can afford a GPU per concurrent stream. For a server-side dubbing pipeline, yes. For a phone app that must work on a plane, no.

Voice design versus a fixed cast

VoxCPM2's most distinctive feature is voice design: a parenthetical description in the text prompt produces a voice with no reference audio at all. Add controllable cloning, where a short clip sets timbre and a style instruction sets delivery, and you have a system for producing a cast of characters rather than reading a script.

Phonon has none of that. Its 37 English voices are a fixed set, and custom voices require a .safetensors file rather than a sentence of description. That is a deliberate simplification: a voice catalogue is easier to audit, easier to license and impossible to prompt into saying something in a celebrity's voice. For compliance-sensitive deployments, the constraint is a feature. For creative work, it is a wall.

Both are Apache 2.0, which changes the calculus

This is the part that makes the comparison cleaner than Phonon's other matchups. VoxCPM2 is Apache 2.0 for code and weights. Phonon 31M is Apache 2.0 for weights, with the runtime under MIT or Apache-2.0. Unlike Gradium's 40M checkpoints, which are CC BY-NC 4.0 and unusable in commercial products without an agreement, both of these can ship.

The licensing difference inside Gradium's own family is worth repeating because it is easy to miss: the model that sounds best — Phonon 40M, at an Elo of 2,117 versus 2,098 for the 31M — is the one you cannot use commercially. The Apache 2.0 model is the smaller, English-only one.

Where each one wins

Choose VoxCPM2 when you need languages beyond English, 48 kHz output, voice design, cloning, or long-form narration, and when a GPU is already part of the deployment. It is also the better choice for research on multilingual TTS, because the checkpoint, the technical report and the evaluation set are all public.

Choose Phonon 31M when the audio must be generated on the device, when English is enough, when the voice set can be fixed, and when a 64 MB download is the difference between shipping and not shipping. Games, accessibility tools, kiosks, browser features and offline assistants are its territory. A 2B model cannot follow it there, and it is not trying to.

OrcaRouter does not host either model — VoxCPM2 is served through vLLM-Omni or a local runtime, and Phonon runs on your hardware. The routing layer is relevant one step up the stack, where the script for the voice is generated: more than 200 language models behind one key, provider list price passed through at 0% markup, and automatic failover when a provider falters. The TTS stays where it belongs; the text generation can move.

What the next benchmark should settle

The missing experiment is obvious and nobody has run it: both checkpoints, one test set, one transcription model, one scorer, published with the audio. Until that exists, the honest position is that VoxCPM2 is the more capable system and Phonon 31M is the more deployable one, and the 0.51-versus-1.84 WER comparison is a coincidence of methodology rather than a result.

What is not in doubt is the direction of travel. A 64 MB English model in October 2026 is credible for production read-aloud work in a way that a 64 MB model would not have been in early 2025. The interesting question for the next release cycle is not whether a small model can match a large one on English. It is whether the small model can pick up a second language without doubling in size. Gradium's 40M multilingual checkpoint is a first attempt, and its non-commercial licence suggests the company has not yet decided how to sell it.