The average is flattering. The Thai score is not.
Edge0 has released Audio8-TTS-Realistic, a compact open-weight text-to-speech model that clones a voice from one reference clip and an exact transcript. On the company’s own nine-language scoreboard, it reports 4.86% average error. Hard subsets (more difficult prompts) are left out. That number will travel. It should not travel alone.
In the per-language breakdown Edge0 published with the release, Thai word error is 48.57%. Finnish is 32.28%. Vietnamese is 28.11%. None of those three sits inside the 4.86%. A headline can look competitive and still be unusable in the market you actually wanted. The averages are not a deployment plan.
What actually shipped
Realistic is the second 0.6B model in the Audio8 family, after Audio8-TTS-Preview-0.6b. The weights are Edge0/Audio8-TTS-Realistic-0.6b. Edge0 ships weights and code under Apache 2.0, with a paper and a public sample page.
Edge0’s paper puts the main model at about 601 million parameters, audio codec excluded. Cloning is zero-shot: one clean reference clip, plus a transcript that matches it exactly. The same quick start can also generate speech with no reference voice at all. That second mode is the safer way to try a company voice, and it comes back later.
Coverage is 23 languages, twelve more than Preview. The additions are Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, and Vietnamese.
Count the tables before you quote any of them. They do not share a denominator. The per-language CV3 error breakdown has 22 languages, 500 items each. The one missing from that list is Cantonese, which Realistic inherits from Preview and does support. The headline 4.86% averages only nine of the 22: Chinese, English, Japanese, Korean, German, French, Spanish, Italian, and Russian. When the paper talks about 13 additional languages, it means that 22-language table minus those nine. Cantonese is in neither group. Call that bookkeeping. It is not a hidden language.
MiniMax is a fourth cut. Edge0’s README says that average covers 24 evaluated languages and includes languages outside the supported list, without naming the extras. A figure from that set is a stress test. It is not a promise about the 23.
The published quick start expects a CUDA GPU. Edge0’s README says there is no public ONNX export of Realistic, so a paper figure of 1.1 to 1.2 GiB peak memory on an Apple M2, using INT4 ONNX, cannot be checked against a file you can download.
Read the scoreboard the way a buyer should
The headline results in Edge0’s README are the company’s own. They are not a third-party audit. Edge0 is also explicit that word error and speaker similarity do not measure whether speech sounds human. Lower error is better. Similarity runs from 0 to 1, and higher is a closer voice match. Customers do not experience a spelling test. They experience pace, pauses, and whether a person seems to be on the line.
Three averages, three different tests:
- Seed-TTS, English and Chinese, excluding hard Chinese: 1.25% error, similarity 0.704. This is a mixed score. It averages English word error with Chinese character error.
- CV3, the nine languages above: 4.86% error, similarity 0.661. Hard subsets are out. Even this average mixes units. Chinese, Japanese, and Korean are character error. The other six are word error.
- MiniMax, 24 evaluated languages: 5.89% error, similarity 0.754. Edge0 includes languages outside the supported list. Do not read 5.89% as the score for the 23.
Against its own predecessor, the comparison Edge0 actually published is mixed. On the eight CV3 languages both models share, Russian excluded, error moves from 4.80% for Preview to 4.56% for Realistic, with Realistic similarity at 0.663. Preview similarity on that slice was not reported. On Seed-TTS, the mixed English-word and Chinese-character score ticks from 1.23% to 1.25%, while similarity rises from 0.682 to 0.704. Edge0 says the two releases were evaluated separately, and that Preview similarity was converted from a percentage on its model card onto a 0-1 scale. Accuracy did not improve everywhere.
Against other models, Edge0 calls its table a reference comparison, because pipelines and text normalization differ. On the nine-language CV3 average, Edge0’s 4.86% sits ahead of cited figures for VoxCPM2 at 5.35%, CosyVoice3-1.5B at 6.89%, and Qwen3-TTS-1.7B at 7.73%. Those competitor numbers were not measured in Edge0’s lab. The Higgs and Qwen3-TTS figures in that comparison come from a Boson AI evaluation. The VoxCPM2 and CosyVoice3 figures come from the VoxCPM repository.
Do not treat 7.73% as proof that a 0.6B model beat a 1.7B model. The same comparison cites Qwen3-TTS at 1.01% on Seed-TTS, that mixed two-language score, with hard Chinese left out. A jump from 1.01% to 7.73% is not a small normalization footnote. It is a different test, from a different lab’s table, on a nine-language average rather than an English-Chinese one. Borrowing it for a size victory overstates what the tables can support.
The lead also does not survive the bigger model. Higgs Audio V3, Boson AI’s roughly 4B system, published as bosonai/higgs-audio-v3-tts-4b, is cited at 4.41% on CV3 and 2.74% on MiniMax. On the mixed Seed-TTS score, the reported order is Qwen3-TTS at 1.01%, Higgs Audio V3 at 1.11%, Audio8 at 1.25%, then VoxCPM2 at 1.41% and CosyVoice3 at 1.67%. MiniMax punishes narrow coverage in the other direction too: the same comparison lists Qwen3-TTS-1.7B at 27.41% there and notes that it officially supports only 10 of the 24 test languages. That 27.41% does not judge the languages Qwen claims. The same rule applies to Audio8’s 5.89%.
Inside Edge0’s per-language breakdown, the low-error cluster is Chinese at 1.81%, English at 2.91%, German at 3.20%, and Spanish at 3.39%. Korean is 4.80%. Italian is 4.93%. French is 7.94%, Japanese 7.51%, Russian 7.26%. Then the floor drops: Arabic 21.10%, Hindi 21.71%, and the three high-error languages from the opening, Vietnamese 28.11%, Finnish 32.28%, and Thai 48.57% word error. Those last three are outside the 4.86%. Quoting the headline in a market-entry slide for Bangkok or Helsinki would be a category error.
Cantonese does not belong in that ranking. It is supported, absent from the 22-language table, and cited at 31.35% character error on MiniMax. Character error is not word error. Do not sit that figure next to the Thai score and pretend you have a ranking.
A brand voice that sounds like the right person and says the wrong thing is worse than a voice that obviously fails. Similarity and error are not the same test. Do not let one talk you past the other.
A smaller copy of a bigger idea
Edge0’s paper calls the architecture dual-autoregressive and points to Fish Audio S2 Pro. Fish’s own description of S2 matches the shape. Speech is chopped into frames, and each frame is stored as several codebooks rather than one audio stream. A slow branch walks forward in time and predicts the first codebook. A shallow fast branch fills in the other codebooks in parallel. The expensive sequential work happens once per frame, not once per codebook.
Fish puts that split at about 21 frames a second, with a 4B slow path and a 400M fast path. Edge0’s paper describes a much smaller version of the same split: a slow branch built on Qwen2.5 0.5B, running near 21.5 frames a second, and a shallow fast branch for the remaining codebooks. The codec is excluded from the 601 million count and bundled at 44.1 kHz. If that size claim holds, a pattern a 4B-class system used to chase natural speech has been squeezed toward a model you can contemplate fine-tuning without a special cluster request.
Contemplate is the right verb. The paper lists a context of 2, 048 packed text and audio positions. At 21.5 frames a second, that is a ceiling of about 95 seconds even if every position were an audio frame, and less once text takes seats. This is built for clips. Long reads sit outside the design. Edge0 describes a serving adapter. Speed and cost on your traffic are not in the release. Larger systems in this family have published serving numbers. This one has not. Inexpensive to serve is a hypothesis until someone measures it on your traffic.
The interesting part is the least auditable
The training story is a learned realism critic, and the public account does not let you audit it. Edge0’s paper says the critic was trained on a quality ladder of six open-source TTS systems, ranked in a private blind listening study, with human recordings at the top. The six systems are not named. The study’s size and method are not described. Extra rungs were made by squeezing pauses. Treat the critic as a design claim. It is not a certificate that the audio sounds human.
The gap it is aimed at is real enough. A clip can pass a spelling test and match a timbre and still sound like a GPS. The paper describes three stages. First, multilingual supervised fine-tuning: imitate good examples. Second, a reinforcement pass Edge0 calls GRPO, rewarding fewer word errors and a closer voice match, weighted 0.7 and 0.3 in the paper. Third, another pass guided by the critic, which scores each clip against a matched human recording of the same sentence.
Two constraints are there to stop the obvious cheats. The critic is penalized for recognizing which system made the clip, so it is less likely to reward a generator’s fingerprints than human texture. A duration penalty applies past 1.35 times the reference length, so the model cannot win by talking slowly. The critic head is about 3 million parameters on a frozen WavLM-large encoder. All of that is the paper’s account. It is not an independent replication.
Read the stage path as a path, not a victory lap. On the nine-language CV3 set, Edge0’s paper shows error moving from 5.41% after supervised training, to 5.01% after the word-error and similarity stage, to 4.86% after the realism stage. Similarity on that set ends at 0.661, up from 0.639 after supervised training.
On the 13 additional languages (the 22-language table minus the nine, Cantonese excluded), the middle stage does not help. Error goes from 22.36% after supervised training to 22.53% after the word-error reward, then down to 20.77% after the realism stage. Similarity ends at 0.679, up from 0.638 after supervised training and from 0.637 after the middle stage. The paper says the final stage used only Chinese and English prompts, and still moved those lower-resource aggregates. It did not make them good.
Set the ending figures next to the headline, not next to each other as if they shared a scale. Similarity on those 13 languages, 0.679, sits a bit above the 0.661 on the nine-language set. Error, 20.77%, is more than four times the 4.86%. The voice match can look fine while the words are not. Just over a fifth still come back wrong.
Where it belongs, and where it does not
Used narrowly, this is a practical piece of automation. Internal training clips, product walkthroughs, and localized explainers in Chinese, English, German, or Spanish are the pilot that matches the published error. A support workflow that already handles the text can grow a voice, if a human still checks names, numbers, and anything a customer might act on. The model does not know your refund policy. It says the text you hand it, in a voice you supplied.
The no-reference mode is the cleaner sales experiment: a company voice, on a language where error is low, reading a script a person approved, without cloning anyone. Cloning a sales rep from a scraped call recording is not an experiment. It is an impersonation problem.
Nothing in the public account of this release describes a consent check, a watermark, or a misuse control. That is a statement about the materials, not a claim that no control exists anywhere in the product. A permissive license does not grant rights in someone else’s voice. The rights question belongs on the pilot checklist before the first sample clip.
Edge0 says the model is in TTS Arena and has drawn encouraging ratings. No vote counts or ranks are published. Encouraging is not a score. Listen to the samples, then run a private test on your own scripts. If the critic is doing what the training story claims, you should be able to hear it. If you cannot, the error tables are all you actually have.
Questions worth asking before a pilot
- Does 4.86% mean it beat the larger open models?
Only on Edge0’s nine-language CV3 average, and only inside a reference table whose pipelines do not match: Edge0’s own 4.86%, with Qwen3-TTS and Higgs figures from a Boson AI evaluation and VoxCPM2 and CosyVoice3 figures from the VoxCPM repository. The same comparison cites Qwen3-TTS at 1.01% on a different mixed English/Chinese score, and Higgs Audio V3 at 4.41% on CV3 and 2.74% on MiniMax, so the cited 7.73% is not a clean win for a 0.6B model over a 1.7B one.
- Is it better than the preview model?
On the shared eight-language CV3 slice, error improved from 4.80% to 4.56%. On Seed-TTS, a mixed score of English word error and Chinese character error, similarity rose from 0.682 to 0.704 while error moved from 1.23% to 1.25%, and the two releases were evaluated separately.
- Which of the 23 languages are safe to ship?
None are certified. On Edge0’s table, Chinese, English, German, and Spanish are at 3.39% or below, Korean is 4.80%, Italian is 4.93%, and French, Japanese, and Russian are 7.94%, 7.51%, and 7.26%. Thai at 48.57% word error is not a production voice, and Cantonese’s cited 31.35% is character error on MiniMax, not a number to rank against that Thai figure.
- Can we run it on a laptop this week?
Not on a documented path. The quick start wants a CUDA GPU, and the paper’s 1.1 to 1.2 GiB Apple M2 figure depends on an INT4 ONNX export Edge0 has not published for Realistic.
- Did the realism stage fix the weak languages?
No. The paper says that stage used only Chinese and English prompts, and on the 13 additional languages (the 22-language table minus the nine, Cantonese excluded) error fell from 22.53% to 20.77% while similarity ended at 0.679, up from 0.638 after supervised training. That is a gain, and it is still just over a fifth of the words wrong.
- What stops someone cloning a voice they do not own?
Nothing described in the public account of this release: no consent check, watermark, or misuse control is part of that account, and a permissive license does not grant rights in someone else’s voice. Generating without a reference avoids cloning a person. It does not answer the rights question if the voice you supply is not yours.
Do not let a similarity score talk you into a voice you do not have the right to use.