Turkish is spoken by more than eighty million people, but it has remained a mid-resource language for speech synthesis. A technical report just published by PatientDesk AI puts a 7,620,931-parameter text-to-speech model in front of us: 48 kHz output, running at 41 times real time (RTF 0.024) on an Apple M5 laptop CPU with only two threads, and 30.5 MB of weights. Not one real human recording is in the training data. All of it comes from VoxCPM2, a teacher model roughly 300 times larger that generates synthetic speech, and about 1,790 hours survived automatic quality control.

The interesting part is not the parameters but the almost constant self-questioning in this report.

What decides the outcome is data, not scale

The model card splits the parameters in two: a 3.69M acoustic model that draws mel spectrograms from text, and a 3.93M vocoder that turns those mels back into 48 kHz audio. It has 10.93% fewer parameters than the comparison system EMA Lightning (8,556,172), and EMA Lightning is also trained purely on VoxCPM2 synthetic speech from the same teacher, with a different voice.

Freya-TR-Eval is a public benchmark of 495 sentences. On the model card's own harness: Mini 1.0.1 scores WER 1.02%, EMA Lightning 1.00%, and the teacher VoxCPM2 0.92%. On the public benchmark page's harness instead (other seeds, one NVIDIA L4), Mini posts 1.09% against EMA's 1.04%, and CER 0.220% against 0.225%. The paper calls 0.220% "the lowest of the 11 open-weights systems", yet across all 15 systems it ran it comes second, behind the authors' own closed Alania-2 voice Öncel at 0.217%. The authors' own follow-up sentence is blunter still: these gaps amount to 2 to 7 character errors in 80,922 characters, and "none of them is distinguishable from zero".

The sibling report Alania-2 argues the same point: the 488M version lands within half a point of CER of a 2.3B model, and the 7.6M single-voice model "matches the most intelligible systems we measured".

A vocoder has to be trained on generated mels

The first 2.16M vocoder was trained only on real mel spectrograms, and got worse the longer it trained: copy-synthesis UTMOS on 40 corpus clips fell from 2.54 at step 20,000 to 2.17 at step 148,000, against 3.29 for the teacher. Mel loss fell the whole way down while UTMOS fell the whole way down too, and the two decoupled. The diagnosis is right: at inference the vocoder receives mels generated by the acoustic model, which are not the real mels it trained on.

After fine-tuning on the acoustic model's own outputs instead, word error rate dropped from a mean of 3.89% at adjacent pre-release checkpoints to 1.75%, while a vocoder frozen over the same stretch stayed between 3.63% and 4.14%. The shipped 3.93M vocoder pays about 0.24 UTMOS for that.

The teacher's defect, inherited by the student

Turkish has a distinctive letter, ö. The teacher model VoxCPM2 corrupts the start of a clip when the first word begins with ö, and different words break by different amounts: of the 12,554 kept long-set lines starting with ö (counting both cases), only 26% have that first word right (95% CI 25.4 to 26.9), against 94% for lines starting with other capital letters and 23.66% for capital Ö alone. The same ö word scores about 98% later in a clip, which ties the defect to the clip start.

The student learned it. Quality control kept 9,271 lines whose first ö word Whisper judged wrong, and filtered fine-tunes showed no gain. The paper's stopgap is telling: prepend a spoken "ee," to any chunk starting with ö or ı, then cut that audio back out. That is a cover-up, not a fix. The model card says the trick "hides it at the start of each chunk but does not remove it", and the real fix requires regenerating the data at the source.

Three places the paper admits it is weak

First, the final numbers were selected. The paper says so plainly: "Final-settings numbers are optimistic by selection." The doubling rule came from benchmark errors, and temperature, steps, speed, suffix and EQ were all screened on the same 495 sentences; the public harness only changes the seeds, not the sentences.

Second, no human listening test exists. The paper states "No human listening test exists", and the error analysis adds that nobody has listened to these takes. Every quality number comes from an automatic metric, and the authors concede WER may flatter the model relative to human intelligibility by an amount they have not measured.

Third, its errors repeat more than the competitor's. On the public harness, 70.3% of the model's word errors sit in sentences that fail on all three takes, against 54.9% for EMA Lightning. Sixteen sentences fail all three takes for Mini but not for EMA Lightning, twelve of which "look like real sound errors" from the transcripts alone. UTMOS even points in opposite directions: EMA Lightning leads at full band (3.30 against 3.18) while Mini leads in the 8 kHz band (3.05 against 2.88), and the paper writes that it has no measurement of why the sign changes.

Six thousand steps with no gradients

The report also keeps what reads like an incident post-mortem. bf16 autocast caches the weights it casts, and the authors computed shortcut targets under no_grad inside the same autocast region, so the training pass received detached copies and the generator got no gradients. The signature is quiet: flow-matching loss sits flat near 2.03, validation loss on the moving-average weights never moves, and the shortcut loss is 0. For the first 6,000 steps of the main run and 15,000 of the sibling run, the generator was not learning at all; at step 13,000 of the sibling run, the generator's output weights still had zero norm.

The lesson is one line: after the first step, assert that every trainable module has a gradient. Their first launch crashed on an unused-parameter error, and they silenced it with find_unused_parameters=True long before finding the real cause.

What is open, and what is shut

Apache-2.0 covers the weights, inference code, text normalizer, ONNX graphs, page code, and the whole benchmark apparatus: adapters, results.json, plus per-clip transcripts, scores and timing. The paper says every one of the 20 rows in results.json "can be recomputed with zero difference from the per-clip files" without rendering any audio.

Not published: the training code, data generation scripts, training data, ONNX export tools and Android source. You can run it and rescore it, but you cannot reproduce it. The shipped lineage cost 114.5 GPU-hours (36.4 for the acoustic model, 78.0 for the vocoder), and all training runs came to 159.0.

So what

The most valuable thing in this report is not the 7.62M number but its willingness to write "the teacher's voice breaks at the start", "this number was picked" and "nobody did a listening test" into the body. Automatic metrics tell you how many mistakes a model makes on a benchmark; they do not tell you what is hiding inside the teacher's voice, nor that a training run sat idle for six thousand steps.

Data lineage is the most overlooked asset of this era. A model can be small enough to run on a laptop, fast enough to run in a browser, and downloadable in full — but if the training data is closed and the teacher's defect was never traced to its source, "small" merely means the irreproducibility was scaled down too.