Fish Audio Drama 3 vs VoxCPM2: A Leaderboard Against a Datasheet

Fish Audio Drama 3 has an Elo and VoxCPM2 does not, which makes the obvious comparison between them impossible and the interesting one easy to miss. Drama 3, announced on 13 October 2026 and scored by Artificial Analysis since, holds sixth place of 100 on the Provider Voice Arena at 1,278. VoxCPM2 has no row on any of Artificial Analysis' speech boards, and it is absent from both arenas cited on this page. What both labs do publish is Seed-TTS-eval, and there the numbers say something a single ranking could not: on each vendor's own run, Drama 3 needs two passes to beat VoxCPM2's one. The choice between these two models is not really about quality. It is about whether you want a performance engine you rent by the character, or a two-billion-parameter model you can hold.

One has an Elo, the other has a datasheet

The asymmetry is worth stating plainly because most comparison pages quietly paper over it. Drama 3 is on every board that matters: sixth of 100 in the Provider Voice Arena at 1,278, twelfth of 49 in the Controlled Voice Arena at 1,085, first-equal of 31 on English pronunciation robustness at 93.2%. VoxCPM2 appears on none of them.

That absence is not a scandal, and it is not new. VoxCPM2 has been out since spring 2026 — the project's docs date the 2.0 line to March, the repository tags it 2026.04, and the technical report landed in June as arXiv:2606.06928 — and in all that time nobody has put it into a blind listening arena. The likeliest explanation is boring: arena slots are usually filled by vendors who have a hosted endpoint to promote, and OpenBMB does not sell one. VoxCPM2 ships weights on Hugging Face and ModelScope and an OpenAI-compatible /v1/audio/speech server through vLLM-Omni, and that is the whole distribution story.

So the honest position is this: we can compare Drama 3 against a score, or we can compare the two models against each other. We cannot do both at once, and any page that presents a VoxCPM2-versus-Drama 3 Elo is inventing one.

The one axis both of them publish

Seed-TTS-eval is the closest thing this field has to a common yardstick, and both labs report on it. The catch is that neither publishes a full table in the same format, so what follows is two vendors measuring themselves.

• English word error rate — Drama 3 reports 1.54% on a final take and 2.20% on its first take; VoxCPM2 reports 1.84% in a single pass. The finished Drama 3 take is the best of the three numbers and its first take is the worst.

• Chinese error rate — Drama 3 at 0.83% CER on a final take, 1.00% on its first; VoxCPM2 at 0.97% CER. Same pattern, with Drama 3's final take ahead and its first roughly level.

• Speaker similarity — Drama 3 reports 0.640 on English (a 0–1 scale); VoxCPM2 reports 75.3 on English and 79.5 on Chinese, which on the 0–100 convention used elsewhere in its report is 0.753 and 0.795. VoxCPM2 is the stronger cloner on its own numbers.

• Wider stress — VoxCPM2 also publishes CV3-eval error rates across eleven languages (3.65% zh, 5.00% en, 8.55% hard-zh, 9.85% fr, 4.25% it, 5.21% ru) and an internal 30-language average of 1.68%. Drama 3 publishes no equivalent per-language error table, so there is nothing to put beside them.

Take the similarity gap seriously and the word-error gap lightly. Speaker similarity is the metric most likely to survive a harness difference, because it is a cosine distance between embeddings of the same reference audio. Word error rates depend on normalisation, text preprocessing and which reference clips were used, and the two labs were not required to agree on any of that. The direction is credible — VoxCPM2 clones more faithfully, Drama 3 transcribes more accurately after revision — but the magnitudes are not a measurement.

The revision tax nobody puts in the headline

1.54% and 2.20% are the same model on the same test, differing only in whether the take was accepted first time. Fish Audio's technical report is unusually honest about this: it reports both, because multi-take revision is the product. The pitch is that you direct a line in plain language, listen, redirect, and keep the take you want — and the benchmark reflects the keep.

If that is how your pipeline works, the 1.54% is the number you will see. If you are streaming a single pass into a live call, a game dialogue line or a batch render with no human in the loop, you are buying the 2.20% — worse than VoxCPM2's single-pass 1.84%. That inversion is the most useful fact in this comparison and the one neither vendor's table makes obvious. It also prices the workflow: every extra take is another full synthesis, and Drama 3 bills by input characters, so a three-take habit multiplies the invoice by three.

VoxCPM2 has no revision mechanism of this kind. It offers four generation modes — zero-shot, continuation, reference-only, and combined reference plus prompt audio — and its answer to a bad take is to render again with different style tags. Cheaper per attempt, because you are not paying per character, but with no model-level memory of the rejected take.

The hardware floor is the actual decision

This is where the two models stop being substitutes.

• VoxCPM2 footprint — about 8 GB of VRAM, a real-time factor of roughly 0.30 on an RTX 4090 and 0.13 with Nano-VLLM; the project also reports 1.76 at Q8_0 on an Apple M4 Pro through llama.cpp-omni. It runs on a workstation.

• Drama 3 footprint — Fish Audio publishes no VRAM floor. Its throughput table was measured on H200s: 166.7 frames per second and an 88.9 ms median first-audio at one concurrent stream, 927.6 frames/s at sixteen, 1,687 frames/s at sixty-four, with an RTF of 0.131 at the single-stream setting. Reviewers put the bf16 requirement at a 48 GB card, and community 4-bit builds claim roughly 32 GB. It runs on a server or on somebody else's.

• Throughput caveat — the launch post quotes 271 frames per second from a single H200, while the model card's streaming table tops out at 166.7 at one request. Both can be true of different measurement setups, but it means the headline speed figure is not the one your load test will reproduce.

The arithmetic follows. A million characters is roughly fifteen to twenty hours of finished speech at ordinary narration pace. On a 4090 at an RTF of 0.30 that is five to six hours of GPU time; on Fish Audio's API it is US$10 during the launch promotion and US$50 after it ends on 12 November. Self-hosting VoxCPM2 is cheaper per hour of audio and far cheaper per hour of experimentation — you can iterate on a voice design for an afternoon without watching a meter.

Routing does not dissolve that choice, so it is worth being precise about what it does. OrcaRouter does not host either model; Drama 3 is available through Fish Audio's own API and several third-party platforms, and VoxCPM2 is available as weights you serve yourself. What we do carry is the speech side of the pipeline around them — Gemini's TTS line and OpenAI's TTS-1-HD, with automatic failover across providers, so a hosted endpoint that goes unhealthy does not take the product down with it. If the question is "which of these two do I run", the answer is settled by VRAM, not by routing.

Directing a performance versus designing a voice

The two models disagree about where the creative work happens, and that disagreement is architectural rather than cosmetic.

Drama 3 is a director's tool. You write instructions in square brackets before the words they govern — [choked up, crying quietly] — and change them mid-sentence. Paired tags scope a direction to specific words: <whisper>, <emphasis>, <soft>, <fast>, <slow>, <stress>, <forceful>. Pronunciation can be pinned phoneme by phoneme with CMU Arpabet, pinyin or romaji, and scenes take one voice ID per speaker as an array. Multiple takes can be revised with reference to earlier ones, and word-level inpainting re-renders a single marked word to fit the take around it — though both interactive editing and inpainting are still in invite-only testing, with an editing API listed as forthcoming.

VoxCPM2 is a design tool. Its headline feature is Voice Design: describe a speaker in natural language and get a voice with no reference audio at all. Beyond that it does zero-shot cloning, continuation from a prompt, reference-only cloning with an isolated reference channel (so you do not need the transcript that matches the clip), and style control through tags. The limitation is scope: it shapes a voice and a delivery, not a scene. There is no speaker-array syntax, no multi-character dialogue mode, no word-level repair.

For a two-hander audio drama, that difference decides it before any benchmark does. For a product that needs one consistent brand voice across a catalogue, VoxCPM2's design mode is the shorter path and Drama 3's direction syntax is mostly overhead.

Sample rate, languages, and the codec argument

• Output — VoxCPM2 emits 48 kHz through an asymmetric AudioVAE V2 that encodes at 16 kHz and decodes at 48 kHz; Drama 3 stays on 44.1 kHz, generating through a ten-codebook Descript Audio Codec at about 21.5 frames per second. VoxCPM2's extra 3.9 kHz is real but rarely audible next to a compressed delivery format.

• Architecture — VoxCPM2 is tokenizer-free by design, predicting continuous latents directly on a MiniCPM-4 backbone and arguing that the audio codec is the bottleneck; Drama 3 keeps discrete tokens and spends its 35B on a codebook multi-token-prediction decoder that predicts the nine residual codebooks after the first two. Two bets on where the quality ceiling sits.

• Languages — 83 claimed for Drama 3, against 30 for VoxCPM2 plus nine Chinese dialects (Sichuanese, Cantonese, Wu, Northeastern, Henan, Shaanxi, Shandong, Tianjin, Minnan). If you need Thai and Vietnamese and Hebrew and Swahili, VoxCPM2 has all four and Drama 3 is not obviously better at any of them; if you need 83, only one of these is in the conversation.

• Training data — VoxCPM2 was trained on 2.36 million hours of speech (1.8M Chinese and English plus 560K multilingual), a figure its docs publish precisely. Fish Audio does not publish an equivalent, which is not a criticism so much as another place where the two papers are not written to be compared.

The licence: the one place VoxCPM2 wins cleanly

VoxCPM2 is Apache 2.0. Any company, at any revenue, can embed it, modify it, fine-tune it, redistribute it and sell the output, with no threshold and no negotiation. LoRA and SFT fine-tuning are supported and documented.

Drama 3 is under the Fish Audio Community License v1.0: free for research and non-commercial work, free for commercial use while you and your affiliates stay under US$500,000 in annual revenue, and a negotiated licence above that — or at any size if you want to offer Drama 3 as a TTS, voice-cloning or inference service. Finished client audio is explicitly allowed at any revenue; a competing synthesis API is not. Artificial Analysis flags it as "open weights without commercial-use rights", which is a fair one-line summary of the gate.

There is also a detail to check rather than trust: the Hugging Face metadata for fishaudio/drama-3 still reads fish-audio-research-license, while the card text describes the Community License. Read the licence file in the repository before you build a product on the tag.

For a funded company shipping a voice feature, that is the whole comparison in one line: VoxCPM2 costs you nothing in legal review, Drama 3 costs you a conversation with a sales team.

Pick by constraint, not by score

You have one GPU and a product to ship. VoxCPM2. Eight gigabytes, Apache 2.0, 48 kHz, and a voice design mode that gets you a usable brand voice on the first afternoon. You give up scene direction and the best provider-voice score in the category.

You are producing scripted audio with a human in the loop. Drama 3. Bracketed direction, scoped tags, mid-sentence shifts and word-level repair are the closest thing open weights have to a recording session, and the 1.54% English WER is what the revision loop buys. Budget for the 12 November price change and read the service clause first.

You need a language outside the thirty. Drama 3, or neither — but check the specific language against Fish Audio's own coverage claims rather than the "83" on the tin.

You are cloning a specific person's voice at scale. VoxCPM2's speaker similarity is the best number on this page, and its isolated reference channel means you can clone from a clip without matching transcript. Drama 3's advantage is direction, not fidelity to a reference.

The one thing not to do is rank them. There is no shared arena, no shared human preference study, and no shared error harness — just one model with an Elo and one with a datasheet, and a decision that turns on what hardware you own and what the licence lets you do with it.