GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ihnatov-yuri
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ihnatov-yuri
Model Tree
openai/whisper-large-v3-turbo
Input Modalities
Output Modalities
Supported Functionality
| Path | What |
|---|---|
model.safetensors, config.json, generation_config.json, tokenizer*.json, processor_config.json | PyTorch weights, Hugging Face Transformers format (WhisperForConditionalGeneration), float32 |
coreml/whisper-large-v3-turbo-uk/ | Core ML for Apple devices in WhisperKit's layout, fp16: MelSpectrogram, AudioEncoder (1.2 GB) and TextDecoder (330 MB) .mlmodelc, plus the tokenizer and generation config. Recommended. |
The Core ML folder's config.json names the version ("version": "whisper_b1"). It was converted with Argmax's
whisperkittools, without palettization.
Transformers:
python
from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="ihnatov-yuri/whisper-large-v3-turbo-uk")print(asr("clip.wav", generate_kwargs={"language": "uk", "task": "transcribe"})["text"])WhisperKit (Swift), with the Core ML folder downloaded:
swift
let config = WhisperKitConfig(modelFolder: ".../coreml/whisper-large-v3-turbo-uk", download: false)let pipe = try await WhisperKit(config)let options = DecodingOptions(task: .transcribe, language: "uk")let text = try await pipe.transcribe(audioPath: "meeting.m4a", decodeOptions: options).map(\.text).joined()Set the language to Ukrainian: the model was trained with the language token given, and language detection was not
evaluated. To pin a version, pass revision= (Transformers) or download a fixed commit.
Run in PyTorch with Transformers' default (greedy) generation, the same weights score 8.97% on the mean of the 10 Ukrainian test sets instead of 8.23%, mainly because of repetition on the podcast set (15.39%). WhisperKit's temperature fallback avoids that, so the Core ML build is the one to use.
Speed and size. About as fast as stock turbo, which is several times faster than large-v3, because the decoder has 4 layers instead of 32. Core ML: 1.6 GB.
This section is the same on the cards of Parakeet UK and Whisper UK, so the numbers can be compared directly. Updated 30 October 2026 for Parakeet UK version 3.
Held-out test sets. The same clips for every model: a 1-hour sample of each test set. All models run in Core ML on an Apple M3 Pro: Whisper through WhisperKit with the language set (Ukrainian, or English for the English sets), Parakeet through FluidAudio with no language hint. Parakeet v3 (base) is FluidAudio's own 6-bit build, Parakeet UK v3 the int8 build from its repository and Whisper UK v1 the fp16 build from its repository. WER is on normalised text (lower case, no punctuation, hyphens as spaces). No meeting, session, channel, video or speaker in these sets was used for training.
| Test set | Clips | Whisper large-v3 | Whisper large-v3 turbo | Whisper UK v1 | Parakeet v3 (base) | Parakeet UK v3 |
|---|---|---|---|---|---|---|
| Parliamentary committee meetings | 325 | 12.26% | 13.69% | 9.33% | 14.31% | 8.28% |
| Podcasts and interviews | 269 | 11.59% | 13.83% | 10.75% | 14.17% | 10.62% |
| YODAS, public videos | 459 | 14.56% | 16.99% | 14.35% | 16.17% | 13.20% |
| Verkhovna Rada plenary | 306 | 7.92% | 8.86% | 4.64% | 8.80% | 4.44% |
| Public video talks, mixed topics | 270 | 11.57% | 11.81% | 7.79% | 12.23% | 7.94% |
| FLEURS uk | 343 | 7.37% | 9.00% | 7.01% | 7.20% | 6.36% |
| IT | 270 | 7.58% | 8.14% | 8.47% | 9.30% | 8.54% |
| Medicine | 269 | 11.34% | 11.78% | 7.86% | 10.98% | 8.84% |
| Psychology | 272 | 4.38% | 5.04% | 4.42% | 5.98% | 4.38% |
| Engineering | 248 | 9.36% | – | 7.69% | 11.20% | 8.25% |
| Mean of the 10 Ukrainian sets | 9.79% | – | 8.23% | 11.03% | 8.08% | |
| FLEURS en | 378 | 4.83% | 5.06% | 5.28% | 5.90% | 5.47% |
| AMI English meetings | 862 | 22.30% | 22.63% | 13.21% | 14.14% | 12.37% |
| Code-switching: 200 clips with an English sentence between two Ukrainian ones | Whisper large-v3 | Whisper large-v3 turbo | Whisper UK v1 | Parakeet v3 (base) | Parakeet UK v3 |
|---|---|---|---|---|---|
| WER | 43.17% | 44.48% | 38.98% | 30.95% | 12.09% |
| English words written in English | 0.3% | 2.3% | 9.5% | 32.8% | 78.7% |
Stock turbo did not finish the engineering set, so it has no mean. Pooled over the 10 Ukrainian sets, difference in WER with a 95% interval from a paired bootstrap over recordings (negative = the first model is better):
So on these test sets the two fine-tuned models are about equally accurate, and both are clearly better than Whisper large-v3 and their base models. Parakeet UK keeps English inside Ukrainian speech; Whisper with the language set to Ukrainian mostly drops or transliterates it. Parakeet UK v3 was trained on clips joined the same way as the code-switching test, which flatters it there.
Real online meetings (private test). 35 minutes of Ukrainian one-to-one online meetings, in 8 stretches with each side of the call read separately (4,155 words), run through a Mac transcription app (Transcriberr) with its own audio handling. "Combined" means two engines whose readings are merged word by word. The recordings are not published. The reference transcripts were built blind (engines hidden behind letters) by an AI judge from four engines' readings and were not checked by a person, so treat these numbers as indicative. Parakeet UK v2 was among the judged readings; Parakeet UK v3 and Whisper UK v1 were not, which favours v2. Spoken verb forms ("внідрять" for "внідряти") and "100%" for "100 відсотків" count as correct.
| In the app, 8 stretches | Local microphone | Other side of the call | All | Time |
|---|---|---|---|---|
| Whisper large-v3 | 27.2% | 22.8% | 25.3% | 34 min |
| Parakeet v3 (base) | 26.2% | 18.6% | 23.0% | – |
| Whisper UK v1 | 28.3% | 15.7% | 23.1% | 8 min |
| Parakeet UK v2 | 19.3% | 10.4% | 15.6% | – |
| Parakeet UK v3 | 20.6% | 10.2% | 16.3% | 2 min |
| Combined: Whisper large-v3 + Parakeet UK v2 | 20.6% | 11.9% | 17.0% | 56 min |
| Combined: Whisper UK v1 + Parakeet UK v2 | 16.2% | 11.0% | 14.0% | 9 min |
| Combined: Whisper UK v1 + Parakeet UK v3 | 16.9% | 10.4% | 14.2% | 9 min |
Alone, Parakeet UK is the more accurate of the two fine-tuned models on these meetings; Parakeet alone reads all 8 stretches in about 2 minutes. Combined with Whisper UK v1, versions 2 and 3 of Parakeet UK score the same (difference +0.2 points, 95% interval -0.8 to +1.0), and either beats the large-v3 combination: with v3, by 2.8 points (95% interval 1.2 to 4.4), better on 7 of the 8 stretches.
1 epoch over 617 hours (106,419 windows of up to 30 s, built by packing consecutive clips), lr 3e-5, batch 8 with
4-step gradient accumulation, on a MacBook Pro M3 Pro (MPS). The encoder (32 layers) was frozen and the decoder trained:
171.9 M of 808.9 M parameters. Labels: <|startoftranscript|><|uk|><|transcribe|><|notimestamps|> text, with <|en|>
for English. Validation WER went from 8.07% to 7.44%.
| Source | Hours | Licence of the transcripts |
|---|---|---|
| Verkhovna Rada committee meetings, aligned to the official committee stenograms | 245 | CC BY 4.0 (stenograms) |
| Verkhovna Rada plenary sessions, aligned to the official stenograms | 35 | CC BY 4.0 |
| YODAS Ukrainian (public videos), each video checked against a second recogniser | 120 | CC BY 3.0 |
| Public videos and podcasts with uploader-made Ukrainian subtitles | 113 | see below |
| Mozilla Common Voice 22, Ukrainian | 50 | CC0 |
| FLEURS Ukrainian, train split | 9 | CC BY 4.0 |
| English: AMI meeting corpus 21.8 h, Common Voice English 15 h, FLEURS English 8 h | 45 | CC BY 4.0 / CC0 / CC BY 4.0 |
The same checks as for Parakeet UK apply: machine-made or translated subtitle tracks were rejected, automatically dubbed audio was removed, and test meetings, channels and videos were excluded from training across all sources. AMI's labels are a recogniser's reading, kept only where it agrees with the AMI transcript (at most 15% WER).
A second version, trained to imitate a larger model's output (distillation), scored worse: 9.45% against this version's 8.97% on the mean of the 10 Ukrainian sets, both in PyTorch. It is not published.
MIT, as the base model. Please credit the author (see "How to credit").
Fine-tuned on Verkhovna Rada plenary and committee sessions with the parliament's official stenograms (CC BY 4.0), YODAS (CC BY 3.0, espnet/yodas), Mozilla Common Voice (CC0), Google FLEURS (CC BY 4.0), the AMI meeting corpus (CC BY 4.0), and public videos and podcasts from a video platform with uploader-made Ukrainian subtitles, used for research. That material belongs to its owners; no audio or transcripts are distributed here.
The model is provided as is, without warranty. Whether your use, especially commercial use, is appropriate is your decision and your responsibility. Content owners can ask to be excluded from future versions via https://ihnatov.nl.
Yuri Ihnatov (https://ihnatov.nl), fine-tuned from OpenAI Whisper large-v3 turbo (MIT).
from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="ihnatov-yuri/whisper-large-v3-turbo-uk")print(asr("clip.wav", generate_kwargs={"language": "uk", "task": "transcribe"})["text"])let config = WhisperKitConfig(modelFolder: ".../coreml/whisper-large-v3-turbo-uk", download: false)let pipe = try await WhisperKit(config)let options = DecodingOptions(task: .transcribe, language: "uk")let text = try await pipe.transcribe(audioPath: "meeting.m4a", decodeOptions: options).map(\.text).joined()