GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Slavik2009
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Slavik2009
Model Tree
openai/whisper-large-v3-turbo
Input Modalities
Output Modalities
Supported Functionality
| Model | Task |
|---|---|
| TurkmenChat-7B | 💬 Conversational LLM |
| whisper-large-v3-turbo-turkmen | 🎙️ Speech-to-text |
Protocol. The eval set is the first 300 clips of the corpus, frozen and identical in every session — the model never trains on them (hard-example mining draws only from the training pool). Base and fine-tuned numbers come from the same clips, same model, same decoder — the only difference is whether the LoRA adapter is active. Decoding: beam search (num_beams=5).
Text normalization. WER/CER use one normalizer applied identically to
hypothesis and reference (Unicode NFC, whitespace, and Turkmen orthography:
clitic particles -da/-de written joined, postposition ýaly written
separate) — the same idea as Whisper's EnglishTextNormalizer. We report both
the normalized WER (headline) and the verbatim WER (raw decode) so the
effect of normalization is fully transparent.
| Model | WER (normalized), % | CER, % | WER (verbatim), % |
|---|---|---|---|
| Whisper large-v3-turbo, zero-shot | 124.61 | 90.58 | 124.61 |
| This model (fine-tuned) | 24.44 [95% CI 22.5–26.6] | 5.08 | 24.58 |
| Improvement | −100.2 pts (80% relative) |
Word-level error decomposition on the fixed eval set — shows what kind of errors remain, not just how many:
| Substitutions | Deletions | Insertions |
|---|---|---|
| 21.07% | 2.00% | 1.37% |
Substitutions dominate for this model — mostly close phonetic/orthographic letter confusions (voicing d/t, b/p, g/k) rather than dropped or invented words, which is the expected failure mode for a well-aligned ASR system.
| Session | WER, % | CER, % |
|---|---|---|
| 2 | 39.75 | 8.60 |
| 3 | 34.74 | 7.30 |
| 5 | 29.51 | 6.26 |
| 6 | 27.89 | 6.23 |
| 7 | 24.44 | 5.08 |
Best WER so far: 24.44% · Sessions: 7 · Corpus position: 72750 files · This session: 13164 train clips.
| Reference (human) | Base Whisper (zero-shot) | This model |
|---|---|---|
| esalawmalikim hürmeti adamlar garaşsyz baky bitarap ýurdumyzdaky | Assalamualaikum warimata adammar gharas shada, bachibi tarabi yorna muddhaqe. | suluwmalikim hürmetli adamlar garaşşjada baky bitarap ýurdumyzdaky |
| dürdäne sygyrlaryny içgin öwrenýärler | Дұрдайна сығырларын, ечкейна оранярларын, | dürdäne sygyrlaryny içgine öwrenýärler |
| öz dillerine terjime edýärler | Öz dillerinə tercüm etdiərlər, | öz dillerine terjime etýärler |
| çünki arkadagly gahrymançy ýerdarymyzyň | Tшоң кей, арқыда алғы ақыл мансиердармыдың, | çünki arkadagly gahryman serdarymyzyň |
| bu ajaýyp kitabyndade bellenişi ýaly | Bu acayip kıtahmla de bellensiali, |
Full per-clip data (both decodes, all metrics): benchmark_samples.json.
| Base model | openai/whisper-large-v3-turbo (809M params) |
| Adaptation | LoRA (r=32, α=64, dropout=0.05) on q/k/v/o + fc1/fc2 |
| Precision | bf16 (AMP) + gradient checkpointing |
| Optimizer schedule | cosine, warmup 100 steps, LR 5e-5 (continued) |
| Epochs / session | 2 |
| Effective batch | ~32 |
| Regularization | SpecAugment (time/feature masking p=0.05) |
| Hard-example mining (OHEM) | up to 250 training clips in the WER band [0.15,0.85], upweighted ×2–8 ∝ error severity (substitution-heavy clips boosted) |
Source: mamed0v/TurkmenSpeech. Every clip passes a quality gate before training: duration 1.0–30.0s · chars-per-second 2.0–35.0 (catches audio/text misalignment) · RMS silence filter · peak loudness normalization · Unicode NFC + orthographic normalization · transcript de-duplication.
Each session, before training, the current model transcribes a random sample of up to 4000 clips from the training pool and scores per-clip WER. Selection is deliberately principled rather than "just take the worst":
[0.15, 0.85] WER — clips
the model already gets right (nothing to learn) and clips it gets almost
entirely wrong (usually mislabeled/noisy audio, not genuine difficulty) are
both excluded. Upweighting the latter would amplify label noise
(confirmation bias), which is the classic failure mode of self-training.The eval set is never mined, so the benchmark stays honest.
python
from transformers import pipelineasr = pipeline("automatic-speech-recognition", model="Slavik2009/whisper-large-v3-turbo-turkmen")print(asr("audio.mp3"))
Transcribes Latin-script spoken Turkmen (broadcast/podcast-style audio, 1–30s clips at 16 kHz). Quality degrades on heavy dialects, far-field or noisy audio, and code-switched speech. The eval set is an in-domain held-out slice of the training corpus, not an independent test set — treat the numbers as an upper bound of in-domain quality; expect higher WER on out-of-domain audio. Reported CIs are bootstrap estimates over 300 clips and quantify eval-set sampling noise only, not distribution shift.
Fixed seed (42), deterministic corpus ordering, and a public train_state.json
(session count, corpus position, per-corpus indices, benchmark) let anyone
resume or audit the exact training trajectory.
bibtex
@misc{whisperturbotk,title = {whisper-large-v3-turbo-turkmen: Whisper large-v3-turbo fine-tuned for Turkmen ASR},author = {Slavik2009},year = {2026},url = {https://huggingface.co/Slavik2009/whisper-large-v3-turbo-turkmen}}
| bu ajaýyp kitabyndada bellenişi ýaly |
| dünýä deýasynyň ägirdi | Dunjai deia tenneng eigerdi. | dünýä deýatynyň ägirdi |
| Eval decoding | beam search (num_beams=5) |
| Hardware | 1× NVIDIA A100-SXM4-40GB |
| Seed | 42 (deterministic corpus ordering) |