GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
eek
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
eek
Model Tree
openai/whisper-large-v3-turbo
Input Modalities
Output Modalities
Supported Functionality
sh
hf download eek/whisper-turbo-ro-en transcribe.py requirements.txt requirements-speakers.txt --local-dir vorbacd vorbapip install -r requirements.txt # transcripts (faster-whisper)pip install -r requirements-speakers.txt # only for --speakers (PyTorch + NVIDIA NeMo) python transcribe.py interview.m4a # plain transcriptpython transcribe.py interview.m4a --speakers # [00:01.20] S1: ...python transcribe.py interview.m4a --speakers --srt out.srt --json out.jsonpython transcribe.py clip.mp4 --language auto --mode cover # detect language; dense speechtranscribe.py reads any audio or video file FFmpeg can decode, runs on a GPU when one is
available (otherwise on the CPU, slower), and downloads the models on first use. Options:
| Option | Meaning |
|---|---|
--language ro|en|auto | force the language (default ro; Romanian also handles Romglish) or detect it |
--mode native (default) | Whisper's own 30 s seeking. Best for far-field speech, children, conversations |
--mode cover | windows that end in pauses, each prompted with the previous window's text. Best for clean, dense speech (videos, podcasts), and silent on music |
--speakers | add the diarization model; each word gets a speaker (needs NVIDIA NeMo, see requirements) |
--json, --srt | write words with times and speakers, or subtitles |
The folder ctranslate2-int8/ is a CTranslate2 int8 conversion.
python
from faster_whisper import WhisperModelfrom huggingface_hub import snapshot_download path = snapshot_download("eek/whisper-turbo-ro-en", allow_patterns=["ctranslate2-int8/*"]) + "/ctranslate2-int8"model = WhisperModel(path, device="cuda", compute_type="int8_float16") # CPU: device="cpu", compute_type="int8"segments, info = model.transcribe( "audio.wav", language="ro", beam_size=5, condition_on_previous_text=False, word_timestamps=True)for s in segments: print(f"[{s.start:6.2f} → {s.end:6.2f}] {s.text}")Keep condition_on_previous_text=False: Whisper's default of prompting each window with the
previous text makes this model (like most fine-tunes) loop or skip on long recordings.
The repository root holds the model in Hugging Face format (fp16 safetensors).
python
import torchfrom transformers import pipeline asr = pipeline("automatic-speech-recognition", model="eek/whisper-turbo-ro-en", dtype=torch.float16, device="cuda:0")print(asr("clip.wav", generate_kwargs={"language": "ro", "task": "transcribe"})["text"])For audio longer than 30 seconds prefer faster-whisper or transcribe.py.
acuma, aicea, anu), numbers usually as digits.[S1] … [S2] … marks speaker changes inside one 30 s window only. The labels restart in every
window, so strip them (re.sub(r"\[S\d+\]\s*", "", text)) or use the diarization model for
recording-level speakers. transcribe.py does the latter.[, [ and ] tokens (Whisper suppresses
them by default); keep it if you copy the model elsewhere, or the tags cannot be produced.suppress_blank=False (as transcribe.py --mode cover decodes) the model returns nothing
on music and silence instead of Whisper's classic "Mulțumesc pentru vizionare!".Word error rate (%), lower is better, normalized with Gladia's LLM_MARKDOWN_PROTECTED_25 preset plus a Romanian plug-in (diacritics folded, colloquial spellings kept distinct, spoken numbers as digits).
| # | Model | Avg. (9 RO sets) | FLEURS ro | FLEURS en | Films | Spontaneous | Common Voice | Household | ROMPAR | VoxPopuli | Podcasts | Audiobooks | Stories |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | This model | 7.9 | 5.8 | 5.5 | 13.8 | 13.3 | 2.5 | 15.7 | 9.8 | 7.2 | 8.3 | 4.9 | 5.5 |
| 2 | Previous version of this model (ft-w6) | 8.3 | 6.1 | 5.4 | 13.8 | 13.6 | 2.7 | 17.5 | 10.5 | 7.8 | 8.4 | 5.1 | 6.3 |
| 3 | Our Parakeet TDT 0.6B v3 fine-tune + 5-gram LM (unreleased) | 8.7 | 5.9 | 5.4 | 16.4 | 14.4 | 2.2 | 21.4 | 10.0 | 7.5 | 8.1 | 6.5 | 7.2 |
| 4 | surogate/jackrabbit-110m-ro, CTC + 4-gram LM | 9.7 | 6.1 | 96.8 | 19.8 | 16.5 | 1.9 | 30.3 | 11.2 | 5.7 | 10.4 | 7.1 | 8.4 |
| 5 | Our SpeD 110M fine-tune (unreleased) | 10.7 | 8.9 | 98.3 | 20.7 | 17.0 | 2.6 | 35.8 | 11.7 | 8.0 | 8.6 | 8.5 | 10.3 |
| 6 | surogate/jackrabbit-110m-ro, TDT | 10.9 | 8.2 | 98.1 | 22.3 | 18.2 | 2.0 | 32.5 | 11.6 | 6.3 | 10.4 | 9.1 | 9.8 |
| 7 | openai/whisper-large-v3 | 12.1 | 8.1 | 4.2 | 19.4 | 20.0 | 5.3 | 26.4 | 17.7 | 11.5 | 10.4 | 7.3 | 9.3 |
| 8 | IonGrozea/whisper-large-v3-turbo-ro | 12.1 | 9.3† | 9.5 | 22.6 | 14.2† | 3.2† | 40.6 | 17.8 | 11.5† | 12.1 | 8.5 | 10.0 |
| 9 | nvidia/canary-1b-v2 | 12.7 | 6.5 | 4.7 | 27.0 | 21.1 | 8.2 | 31.4 | 14.6 | 6.3 | 11.3 | 8.9 | 10.1 |
| 10 | BuzzASR/romanian (large-v3) | 12.9 | 9.0 | 4.6 | 23.6 | 19.8 | 4.1† | 29.0 | 17.9 | 10.5 | 10.9 | 8.3 | 12.1 |
| 11 | VMadalina/whisper-large-news-augmented2 | 13.4 | 11.5 | 41.9 | 25.6 | 19.9 | 5.6 | 31.4 | 16.8 | 10.9 | 9.3 | 10.6 | 10.1 |
| 12 | upb-nlp/ro-fast-conformer | 13.6 | 9.4 | 87.2 | 29.5 | 21.4 | 2.6 | 52.6 | 15.6 | 6.9 | 14.3 | 9.7 | 12.7 |
| 13 | SpeD ParakeetRomanian 110M, TDT | 13.8 | 10.0 | 99.3 | 27.3 | 21.7 | 3.2 | 79.2 | 13.9 | 10.4 | 13.4 | 10.9 | 13.3 |
| 14 | openai/whisper-large-v3-turbo | 14.3 | 9.0 | 4.6 | 25.2 | 25.9 | 10.9 | 28.8 | 17.1 | 12.9 | 10.5 | 7.4 | 9.5 |
| 15 | nvidia/parakeet-tdt-0.6b-v3 | 17.2 | 12.4 | 5.1 | 32.9 | 24.8 | 9.8 | 48.9 | 22.3 | 9.5 | 14.7 | 13.4 | 15.1 |
| 16 | TransferRapid/whisper-large-v3-turbo_ro | 18.6 | 17.9 | 12.5 | 31.8 | 27.6 | 0.7† | 63.8 | 25.6 | 18.0 | 18.4 | 13.8 | 13.9 |
| 17 | readerbench/whisper-ro (small) | 20.0 | 16.1 | 97.5 | 39.8 | 29.5 | 1.9 | 46.9 | 25.9 | 12.5 | 19.9 | 17.2 | 17.6 |
| 18 | facebook/mms-1b-all (ron) | 20.9 | 12.7 | 18.1 | 41.6 | 30.6 | 5.5 | 66.4 | 29.1 | 11.3 | 23.9 | 15.5 | 18.0 |
| 19 | VladS159/Whisper_medium_ro | 21.5 | 16.1 | 6.2 | 32.8 | 30.8 | 2.4† | 74.0 | 36.2 | 12.6 | 24.6 | 16.4 | 21.6 |
| 20 | gigant/whisper-medium-romanian | 21.8 | 16.3 | 6.9 | 33.4 | 30.0 | 2.2† | 76.4 | 36.4 | 12.8 | 29.6 | 16.3 | 19.0 |
| 21 | FraPiz/whisper-large-v3-turbo-moldovan-romanian | 21.8 | 14.1 | 9.0 | 35.2 | 23.4 | 6.2 | 195.2 | 44.7 | 11.7 | 32.3 | 14.7 | 13.7 |
| 22 | upb-nlp/ro-wav2vec2 (greedy) | 26.5 | 24.3 | 99.4 | 50.7 | 37.2 | 9.5 | 82.3 | 29.4 | 12.1 | 26.2 | 19.9 | 29.0 |
| 23 | Qwen3-ASR 1.7B | 26.9 | 18.9 | 3.2 | 45.8 | 34.2 | 19.3 | 49.6 | 30.8 | 16.6 | 21.8 | 27.1 | 27.7 |
| 24 | Omnilingual ASR CTC 300M | 30.1 | 25.7 | 18.6 | 51.6 | 38.4 | 21.4 | 65.1 | 36.4 | 15.6 | 25.2 | 26.7 | 29.6 |
| 25 | racai/wav2vec2-base-100k-voxpopuli-romanian | 35.2 | 33.6 | 99.7 | 61.8 | 44.9 | 14.7 | 95.2 | 40.9 | 15.3 | 35.6 | 29.0 | 40.8 |
| 26 | gigant/romanian-wav2vec2 (XLS-R, greedy) | 50.9 | 55.3 | 99.7 | 76.2 | 62.8 | 9.9† | 90.0 | 61.7 | 38.4 | 50.7 | 46.7 | 56.2 |
Every Romanian-capable model we could run, on every public Romanian test set with human transcripts that can be downloaded without an access request, plus the household conversation described below. Greedy decoding except where an n-gram LM is named; the language is forced to Romanian (English for FLEURS en) where a model takes one. "Avg." is the unweighted mean of the nine public Romanian sets (FLEURS en and the household conversation are not in it); the household conversation is decoded in 28 s chunks cut at pauses. Clips longer than 30 s are left out (6 in VoxPopuli, 27 in USPDATRO) because Whisper decodes one 30-second window. Best comparable score in each column in bold. † The model's card says it was trained on this set or on an overlapping Common Voice release, so that score is not comparable.
| Set | Speech | Clips | Hours |
|---|---|---|---|
| FLEURS test | read | 883 | 2.5 |
| RO-N3WS films | Romanian films, conversational | 4,723 | 8.3 |
| USPDATRO ("Spontaneous") | spontaneous and dialectal, from online videos | 2,610 | 3.7 |
| Common Voice 21.0 test (mirror) | read, crowd-sourced | 3,929 | 4.6 |
| VoxPopuli test | European Parliament | 1,377 | 4.2 |
| RO-N3WS podcasts | conversational | 1,723 | 3.4 |
| RO-N3WS audiobooks | read | 2,475 | 4.3 |
| RO-N3WS children's stories | narrated | 2,821 | 4.7 |
| ROMPAR test | Romanian and Moldovan parliament | 1,490 | 1.9 |
"Private" sets are held-out conversational recordings that were never trained on; their references are machine transcripts unless marked human.
| Test set | Reference | Whisper turbo (stock) | This model |
|---|---|---|---|
| FLEURS Romanian test | human | 9.0 | 5.79 |
| FLEURS English test | human | 4.6 | 5.48 |
| Private Romanian conversational set, 19 h | machine | 10.6 | 3.82 |
| Private Romanian set from newer recordings, 10 h | machine | 9.8 | 3.61 |
| Same 19 h set, against independent third-party captions | captions | 16.7 | 12.7 |
| Private English conversational set, 4 h | machine | 3.7 | 2.80 |
| Romanian clips under 3 s | machine | 48.6 | 13.3 |
| English words inside Romanian speech | machine | 26.6 | 10.9 |
| Household conversation: one adult, two children, far-field, 5 min, 28 s chunks | human | 28.8 | 15.7 |
For scale: the machine references and the independent captions disagree with each other by 11.9%, and the best commercial system we tried scores 4.2 on the household conversation.
Household conversation decoded whole with --mode native: 13.5. Whole recordings of 30–600 s,
--mode cover: Romanian 4.00, English 1.71. Speaker-attributed WER (cpWER) on multi-speaker
recordings with the companion diarization model, --mode cover: 15.2 (Romanian) and 15.1
(English). On 132 Romanian and 29 English music-or-silence clips the model outputs 3 and 1
words in total in cover mode.
FLEURS test with the correct language token forced. Each round of Romanian training has eroded these further (mean 54.9 for the previous version). Use stock Whisper for them:
| Language | Whisper turbo (stock) | This model |
|---|---|---|
| German | 4.56 | 22.91 |
| Spanish | 2.86 | 23.23 |
| French | 5.11 | 38.35 |
| Italian | 3.77 | 29.13 |
| Portuguese | 4.16 | 66.23 |
| Polish | 5.95 | 71.67 |
| Hungarian | 15.10 | 73.53 |
| Russian | 5.45 | 177.56 |
| Turkish | 6.97 | 108.06 |
| Ukrainian | 7.45 | 139.05 |
| Mean | 6.1 | 75.0 |
Fine-tuned on a private dataset of Romanian and English speech, mostly conversational, in six stages, each continuing from the previous best (full fine-tune, bf16, AdamW, cosine schedule):
[S#] speaker turns, timestamps on half the windows,
the previous window's text as prompt on half, speed perturbation, SpecAugment, and
no-speech windows with empty targets; lr 5e-6.ft-w6.About 10% English throughout kept English from degrading.
cover mode, English music-only windows still produce a few words now and then.MIT, the same licence as the base model (OpenAI Whisper): use, modify and redistribute it
freely, commercial use included, keeping the notice in LLM_MARKDOWN_PROTECTED_33. This covers the
weights, the CTranslate2 conversion and transcribe.py. The fine-tuning data is private and is
not distributed.
hf download eek/whisper-turbo-ro-en transcribe.py requirements.txt requirements-speakers.txt --local-dir vorbacd vorbapip install -r requirements.txt # transcripts (faster-whisper)pip install -r requirements-speakers.txt # only for --speakers (PyTorch + NVIDIA NeMo) python transcribe.py interview.m4a # plain transcriptpython transcribe.py interview.m4a --speakers # [00:01.20] S1: ...python transcribe.py interview.m4a --speakers --srt out.srt --json out.jsonpython transcribe.py clip.mp4 --language auto --mode cover # detect language; dense speechfrom faster_whisper import WhisperModelfrom huggingface_hub import snapshot_download path = snapshot_download("eek/whisper-turbo-ro-en", allow_patterns=["ctranslate2-int8/*"]) + "/ctranslate2-int8"model = WhisperModel(path, device="cuda", compute_type="int8_float16") # CPU: device="cpu", compute_type="int8"segments, info = model.transcribe( "audio.wav", language="ro", beam_size=5, condition_on_previous_text=False, word_timestamps=True)for s in segments: print(f"[{s.start:6.2f} → {s.end:6.2f}] {s.text}")import torchfrom transformers import pipeline asr = pipeline("automatic-speech-recognition", model="eek/whisper-turbo-ro-en", dtype=torch.float16, device="cuda:0")print(asr("clip.wav", generate_kwargs={"language": "ro", "task": "transcribe"})["text"])