Instructions to use eqlx-ai/whisper-large-v3-turbo-tr-aksan with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use eqlx-ai/whisper-large-v3-turbo-tr-aksan with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="eqlx-ai/whisper-large-v3-turbo-tr-aksan")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("eqlx-ai/whisper-large-v3-turbo-tr-aksan") model = AutoModelForSpeechSeq2Seq.from_pretrained("eqlx-ai/whisper-large-v3-turbo-tr-aksan", device_map="auto") - Notebooks
- Google Colab
- Kaggle
whisper-large-v3-turbo-tr-aksan
Yรถresel aฤฤฑzlarฤฑ daha iyi anlayan Tรผrkรงe konuลma tanฤฑma modeli. Doฤu'dan Ege'ye, Karadeniz'den Trakya'ya yรถresel konuลmalarฤฑ metne รงevirir. Standart Tรผrkรงedeki baลarฤฑsฤฑnฤฑ korur, sessizlikte cรผmle uydurmaz.
Turkish speech recognition tuned for regional dialects. English below.
Tรผrkรงe
Kฤฑsaca
| Eฤitimsiz Whisper | Bu model | Ne oldu | |
|---|---|---|---|
| Yรถresel aฤฤฑzlar | %32,5 hata | %25,8 hata | hatalar beลte bir azaldฤฑ |
| Doฤu aฤzฤฑ | %69,9 hata | %46,8 hata | hatalar รผรงte bir azaldฤฑ |
| Gรผneydoฤu aฤzฤฑ | %41,9 hata | %23,9 hata | hatalar neredeyse yarฤฑya indi |
| Haber ve medya kayฤฑtlarฤฑ | %19,0 hata | %9,3 hata | hatalar yarฤฑya indi |
| Standart Tรผrkรงe | %5,3 hata | %5,3 hata | aynฤฑ kaldฤฑ, hiรงbir ลey kaybedilmedi |
| Sessizlikte cรผmle uydurma | 100 klipte 75 kez | 0 kez | tamamen ortadan kalktฤฑ |
Hata oranฤฑ (WER) nedir?
Konuลma tanฤฑma modelleri kelime hata oranฤฑ (Word Error Rate, WER) ile รถlรงรผlรผr: modelin yazdฤฑฤฤฑ metin, insanฤฑn dinleyip yazdฤฑฤฤฑ doฤru metinle kelime kelime karลฤฑlaลtฤฑrฤฑlฤฑr ve her 100 kelimeden kaรงฤฑnฤฑn hatalฤฑ olduฤu sayฤฑlฤฑr. รรง tรผr hata vardฤฑr:
- Yanlฤฑล kelime: "gidiyoruz" yerine "gidiyoz"
- Eksik kelime: sรถylenen bir kelimenin hiรง yazฤฑlmamasฤฑ
- Fazla kelime: sรถylenmeyen bir kelimenin eklenmesi
รrnek: 10 kelimelik bir cรผmlede model 2 kelimeyi yanlฤฑล yazdฤฑysa WER %20'dir. Dรผลรผk olan iyidir; %0 kusursuz demektir.
Bu kartta iki tรผr deฤiลimden sรถz ediyoruz:
- Puan farkฤฑ: %32,5'ten %25,8'e inmek 6,7 puanlฤฑk bir dรผลรผลtรผr.
- Oransal azalma: aynฤฑ dรผลรผล, hatalarฤฑn yaklaลฤฑk beลte birinin (%21) ortadan kalkmasฤฑ demektir.
Sonuรงlar
1. Yรถresel aฤฤฑzlar: kapalฤฑ test
Modelin hiรง gรถrmediฤi kayฤฑtlardan oluลan bir test seti (holdout) kuruldu: 8 bรถlgeden 92 klip. Bu kayฤฑtlar ve iรงindeki konuลmacฤฑlar eฤitimde hiรง kullanฤฑlmadฤฑ. Test seti, model seรงimi bittikten sonra yalnฤฑzca bir kez aรงฤฑldฤฑ; yani sonuรง, modelin "ezberlediฤi" deฤil gerรงekten รถฤrendiฤi ลeyi gรถsterir.
| Bรถlge | Eฤitimsiz | Bu model | Deฤiลim |
|---|---|---|---|
| Tรผm bรถlgeler | 32,5 | 25,8 | โ6,7 puan (hatalarฤฑn %21'i gitti) |
| Doฤu | 69,9 | 46,8 | โ23,1 puan (%33) |
| Gรผneydoฤu | 41,9 | 23,9 | โ18,0 puan (%43) |
| Marmara | 36,1 | 27,2 | โ8,9 puan (%25) |
| Trakya | 19,8 | 15,0 | โ4,8 puan (%24) |
| Karadeniz | 27,9 | 24,3 | โ3,6 puan (%13) |
| Ege | 16,8 | 14,8 | โ2,1 puan (%12) |
| ฤฐรง Anadolu | 17,6 | 18,3 | +0,8 puan (deฤiลmedi) |
| Akdeniz | 21,1 | 27,5 | +6,3 puan (kรถtรผleลti) |
En bรผyรผk kazanรง, eฤitimsiz modelin en รงok zorlandฤฑฤฤฑ yerlerde: Doฤu ve Gรผneydoฤu. ฤฐรง Anadolu'da fark yok, Akdeniz'de ise sonuรง kรถtรผleลti; bu iki bรถlge daha fazla veriye ihtiyaรง duyuyor (bkz. Amaรง).
ฤฐstatistik ayrฤฑntฤฑlarฤฑ (gรผven aralฤฑklarฤฑ)
Kรถลeli parantez iรงindeki aralฤฑklar %95 gรผven aralฤฑฤฤฑdฤฑr: gerรงek farkฤฑn bรผyรผk olasฤฑlฤฑkla bu aralฤฑkta olduฤunu sรถyler. Aynฤฑ kaynak kaydฤฑn klipleri birlikte yeniden รถrneklenerek (bootstrap) hesaplandฤฑ. Aralฤฑฤฤฑn tamamฤฑ sฤฑfฤฑrฤฑn altฤฑndaysa iyileลme tesadรผf deฤildir.
| Bรถlge | Klip | Fark [%95 aralฤฑk] |
|---|---|---|
| Tรผm bรถlgeler | 92 | โ6,7 [โ14,9; โ1,0] |
| Tekrar dรถngรผsรผne giren 5 klip hariรง | 87 | โ3,0 [โ4,4; โ1,6] (25,2 โ 22,2) |
| Doฤu | 13 | โ23,1 [โ75,5; โ0,4] |
| Gรผneydoฤu | 12 | โ18,0 [โ27,6; โ7,7] |
| Marmara | 8 | โ8,9 [โ11,2; โ5,9] |
| Trakya | 5 | โ4,8 [โ6,0; +4,3] |
| Karadeniz | 16 | โ3,6 [โ6,1; โ1,2] |
| Ege | 12 | โ2,1 [โ4,2; โ0,9] |
| ฤฐรง Anadolu | 14 | +0,8 [โ1,3; +6,4] |
| Akdeniz | 12 | +6,3 [โ2,4; +15,1] |
"Tekrar dรถngรผsรผ": bazฤฑ uzun kliplerde iki model de aynฤฑ kelimeyi defalarca yazdฤฑ. Bu 5 klip รงฤฑkarฤฑldฤฑฤฤฑnda da iyileลme net biรงimde sรผrรผyor.
2. Standart Tรผrkรงe korundu
Az veriyle yapฤฑlan ince ayarฤฑn bilinen riski, modelin eskiden iyi bildiฤi ลeyi unutmasฤฑdฤฑr. Google FLEURS standart Tรผrkรงe testinde (300 klip) hata oranฤฑ eฤitimden รถnce de sonra da %5,3: hiรงbir ลey kaybedilmedi.
3. Haber ve medya kayฤฑtlarฤฑ
Eฤitimde hiรง kullanฤฑlmayan MediaSpeech haber/medya testinde (200 klip) hata oranฤฑ %19,0'dan %9,3'e indi. Model yalnฤฑzca yรถresel aฤฤฑzlarda deฤil, gerรงek hayattan gelen genel kayฤฑtlarda da daha iyi.
4. Sessizlikte cรผmle uydurmuyor
Whisper'ฤฑn bilinen bir sorunu, sessizlik ya da gรผrรผltรผ duyunca "ฤฐzlediฤiniz iรงin teลekkรผrler" gibi hiรง sรถylenmemiล cรผmleler yazmasฤฑdฤฑr. 100 gรผrรผltรผ/sessizlik klibinde:
| Eฤitimsiz | Bu model | |
|---|---|---|
| Bir ลey yazdฤฑฤฤฑ klip | 100 | 12 |
| Akฤฑcฤฑ, uydurma bir cรผmle yazdฤฑฤฤฑ klip | 75 | 0 |
Kalan 12 klipte de cรผmle deฤil, "Tamam" gibi tek bir kฤฑsa kelime yazฤฑyor. Diฤer kliplerde yalnฤฑzca bir nokta koyuyor; aลaฤฤฑdaki kullanฤฑm รถrneฤi bunu boล metin sayar.
Amaรง
Bu model bir yapฤฑlabilirlik รงalฤฑลmasฤฑdฤฑr. Sorusu ลuydu: genel amaรงlฤฑ bir konuลma tanฤฑma modeli, az miktarda yรถresel konuลmayla, standart Tรผrkรงedeki baลarฤฑsฤฑnฤฑ kaybetmeden yรถresel aฤฤฑzlarda iyileลtirilebilir mi?
Cevap evet: yalnฤฑzca 3,4 saatlik etiketli yรถresel konuลmayla (bรถlge baลฤฑna yarฤฑm saat ile bir saat arasฤฑ) hatalar beลte bir azaldฤฑ. Bu bir baลlangฤฑรง noktasฤฑdฤฑr. Daha รงok yรถresel kayฤฑt, daha รงok farklฤฑ konuลmacฤฑ ve รถzenle doฤrulanmฤฑล etiketlerle รงok daha iyi sonuรงlar alฤฑnmasฤฑ beklenir. Ek verinin en รงok fayda saฤlayacaฤฤฑ yerler hรขlรข en zayฤฑf kalan bรถlgelerdir: Doฤu, ฤฐรง Anadolu ve Akdeniz.
Nasฤฑl kullanฤฑlฤฑr
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="eqlx-ai/whisper-large-v3-turbo-tr-aksan",
torch_dtype="auto", device="cuda") # GPU yoksa device="cpu"
out = asr("kayit.wav", chunk_length_s=30, batch_size=8,
generate_kwargs={"language": "turkish", "task": "transcribe"})
text = out["text"].strip()
if not any(ch.isalnum() for ch in text): # sessizlikte model bazen yalnฤฑzca "." yazar
text = ""
Uzun kayฤฑtlarda รถnรผne bir ses etkinliฤi algฤฑlayฤฑcฤฑsฤฑ (VAD, รถr. Silero VAD) koymak รถnerilir: konuลma olmayan bรถlรผmler hiรง modele gitmez.
Model bilgisi. Temel model openai/whisper-large-v3-turbo
ยท 809M parametre ยท fp16, 1,6 GB ยท Lisans CC-BY-NC-4.0 ยท Ek aฤฤฑrlฤฑklar ana modele birleลtirildi, peft gerekmez.
Nasฤฑl eฤitildi
Yรถntem: LoRA ile ince ayar. Ana modelin 815 milyon aฤฤฑrlฤฑฤฤฑ olduฤu gibi bฤฑrakฤฑldฤฑ; yanฤฑna yalnฤฑzca 6,5 milyon (%0,8) yeni, kรผรงรผk aฤฤฑrlฤฑk eklenip bunlar eฤitildi (LoRA, Low-Rank Adaptation). Bu yรถntem az veriyle รงalฤฑลฤฑrken modelin eski bilgisini korur. Eฤitim bitince bu ek aฤฤฑrlฤฑklar ana modelin iรงine birleลtirildi.
Veri: toplam 6,4 saat, 1536 klip.
| Parรงa | Sรผre | Klip | Neden |
|---|---|---|---|
| Yรถresel konuลma (rรถportaj, sohbet) | 3,4 saat | 587 | yรถresel aฤฤฑzlarฤฑ รถฤrenmek |
| Google FLEURS standart Tรผrkรงe (CC-BY-4.0) | 2,5 saat | 749 | standart Tรผrkรงeyi unutmamak |
| รretilmiล gรผrรผltรผ ve sessizlik, boล cevaplฤฑ | 0,5 saat | 200 | sessizlikte cรผmle uydurmamak |
Yรถresel konuลmanฤฑn bรถlgelere daฤฤฑlฤฑmฤฑ (saat): Doฤu 0,90 ยท Ege 0,71 ยท ฤฐรง Anadolu 0,49 ยท Trakya 0,37 ยท Akdeniz 0,34 ยท Karadeniz 0,31 ยท Gรผneydoฤu 0,25. Kayฤฑtlar metne dรถkรผldรผ (etiketlendi); ลรผpheli klipler insan tarafฤฑndan dinlenip dรผzeltildi. Konuลmanฤฑn รงok seyrek olduฤu klipler (mรผzik, uzun sessizlik) รงฤฑkarฤฑldฤฑ. Test kayฤฑtlarฤฑ ve konuลmacฤฑlarฤฑ eฤitimden tamamen ayrฤฑ tutuldu; test cรผmlelerinin hiรงbiri eฤitim metinlerinde geรงmiyor. Ses dosyalarฤฑ bu depoda yayฤฑnlanmaz.
Eฤitim seyri. Kayฤฑp (loss: modelin doฤru metinden ne kadar uzak olduฤunu gรถsteren sayฤฑ) ilk turda hฤฑzla dรผลtรผ, sonra dengelendi.
Eฤitim boyunca her 50 adฤฑmda modelin bir kopyasฤฑ kaydedildi ve ayrฤฑ bir geliลtirme setinde รถlรงรผldรผ. Yรถresel hata adฤฑm adฤฑm %29'dan %20'ye indi; sessizlikte bir ลey yazma oranฤฑ %100'den %12'ye dรผลtรผ. En iyi sonuรง veren son kopya seรงildi.
Teknik ayarlar
| Ayar | Deฤer |
|---|---|
| LoRA rank / alpha / dropout | 32 / 16 / 0,05 |
| LoRA hedefleri | q_proj, v_proj, kodlayฤฑcฤฑ (encoder) ve รงรถzรผcรผde (decoder) |
| Eฤitilen parametre | 6,55M / 815M (%0,80) |
| Optimizasyon | AdamW, aฤฤฑrlฤฑk azaltma 0, gradyan kฤฑrpma 1,0 |
| รฤrenme hฤฑzฤฑ | 1e-4 tepe; ilk %10 (29 adฤฑm) doฤrusal ฤฑsฤฑnma, sonra doฤrusal azalma |
| Etkin batch | 16 (8 ร 2 gradyan biriktirme) |
| Tur / adฤฑm | 3 / 288 |
| Hassasiyet | bf16 autocast, gradient checkpointing |
| Kayฤฑp | token รงapraz entropisi, รถnce her klip iรงinde, sonra klipler arasฤฑnda ortalanฤฑr (kฤฑsa ve boล cevaplฤฑ klipler uzun kliplerle eลit aฤฤฑrlฤฑk alฤฑr) |
| Hedef metin | <|tr|><|transcribe|><|notimestamps|> + metin + <|endoftext|>; en fazla 440 token |
| Klipler | ortalama 21 sn (4โ25 sn); saniyede 1 kelimenin altฤฑndaki klipler รงฤฑkarฤฑldฤฑ |
| Kopya seรงimi | her 50 adฤฑmda bir kopya, geliลtirme setinde (141 klip) WER ile รถlรงรผldรผ; adฤฑm 288 seรงildi |
| Donanฤฑm | tek NVIDIA RTX 4090 (24 GB), en yรผksek bellek 6,7 GB |
Lisans
CC-BY-NC-4.0: araลtฤฑrma ve kiลisel kullanฤฑm serbest, ticari kullanฤฑm yok. Modeli indirip deneyebilir, รผzerinde รงalฤฑลabilir ve paylaลabilirsiniz; paylaลฤฑrken bu depoya kaynak gรถsterin. รcretli bir รผrรผnde veya hizmette kullanmak iรงin izin gerekir. Bu model bir yapฤฑlabilirlik รงalฤฑลmasฤฑ olarak yayฤฑnlanmฤฑลtฤฑr. Temel model Whisper MIT lisanslฤฑdฤฑr; standart Tรผrkรงe verisi Google FLEURS'tรผr (CC-BY-4.0).
Sฤฑnฤฑrlar
- Test setleri kรผรงรผk. Bรถlge baลฤฑna 5โ16 klip; bรถlge bรถlge sonuรงlar kaba bir gรถstergedir.
- Her bรถlgede kazanรง yok. ฤฐรง Anadolu'da deฤiลim yok, Akdeniz'de kรถtรผleลme var; Doฤu'da hata hรขlรข yรผksek.
- Tekrar dรถngรผsรผ. Uzun veya ลarkฤฑya benzeyen kliplerde model ara sฤฑra aynฤฑ kelimeyi tekrar tekrar yazabilir (kapalฤฑ testte 92 klipten 5'i, eฤitimsiz modelde de gรถrรผlรผyor).
- Yazฤฑm stili. Test metinleri eฤitim metinleriyle aynฤฑ sรผreรงten geรงti; kazancฤฑn bir kฤฑsmฤฑ noktalama ve sayฤฑ yazฤฑmฤฑ gibi stile uyumdan gelebilir.
- Kapsam dฤฑลฤฑ. Kรผrtรงe, Zazaca, Arapรงa ve diฤer diller; kimin konuลtuฤunu ayฤฑrma; zaman damgasฤฑ. Tฤฑbbi, hukuki veya kimlik doฤrulama gibi hatanฤฑn zarar vereceฤi kullanฤฑmlar iรงin tek baลฤฑna uygun deฤildir.
English
Turkish speech recognition tuned for regional dialects. Fine-tuned from openai/whisper-large-v3-turbo on 3.4 h
of regional Turkish speech. It keeps its accuracy on standard Turkish and no longer invents sentences on silence.
At a glance
| Base Whisper | This model | Change | |
|---|---|---|---|
| Regional dialects (holdout) | 32.5% WER | 25.8% WER | one fifth of errors gone |
| Eastern | 69.9% | 46.8% | one third fewer errors |
| South-Eastern | 41.9% | 23.9% | almost half the errors |
| News and media (MediaSpeech) | 19.0% | 9.3% | errors halved |
| Standard Turkish (FLEURS) | 5.3% | 5.3% | unchanged |
| Invented sentences on silence | 75 of 100 clips | 0 | gone |
What is WER?
Word error rate: the model's transcript is compared word by word with a human reference, and wrong, missing and extra words are counted per 100 words. 2 wrong words in a 10-word sentence is 20% WER. Lower is better.
Results
Holdout, regional dialects. 92 clips from 8 regions; no recording or speaker was used in training; the set was opened once, after the model was chosen. The per-region chart and table are in the Turkish section.
| Region | Base | This model | Change [95% CI] |
|---|---|---|---|
| All | 32.5 | 25.8 | โ6.7 [โ14.9; โ1.0] |
| All, without 5 clips where either model loops | 25.2 | 22.2 | โ3.0 [โ4.4; โ1.6] |
| Eastern | 69.9 | 46.8 | โ23.1 [โ75.5; โ0.4] |
| South-Eastern | 41.9 | 23.9 | โ18.0 [โ27.6; โ7.7] |
| Marmara | 36.1 | 27.2 | โ8.9 [โ11.2; โ5.9] |
| Thrace | 19.8 | 15.0 | โ4.8 [โ6.0; +4.3] |
| Black Sea | 27.9 | 24.3 | โ3.6 [โ6.1; โ1.2] |
| Aegean | 16.8 | 14.8 | โ2.1 [โ4.2; โ0.9] |
| Central Anatolia | 17.6 | 18.3 | +0.8 [โ1.3; +6.4] |
| Mediterranean | 21.1 | 27.5 | +6.3 [โ2.4; +15.1] |
95% intervals from a bootstrap over source recordings. Other tests: FLEURS tr test (300 clips) 5.3 โ 5.3; MediaSpeech tr (200 clips) 19.0 โ 9.3; noise/silence (100 clips): any output 100 โ 12, fluent invented sentence 75 โ 0.
Purpose
A feasibility study: with only 3.4 hours of labelled regional speech (half an hour to an hour per region), a general-purpose recogniser improves clearly on regional dialects without losing standard Turkish. More regional recordings, more speakers and carefully verified labels should give much better results, most of all in the weakest regions (Eastern, Central Anatolia, Mediterranean).
How to run
Use the Python snippet in the Turkish section: the transformers ASR pipeline with language="turkish" and
chunk_length_s=30; treat a punctuation-only output as empty; put a VAD (e.g. Silero VAD) in front of long recordings.
Training
LoRA fine-tune (6.55M of 815M parameters trained, 0.80%), merged into the base weights; no peft needed. Data: 6.4 h in
1536 clips: 3.4 h regional speech (587 clips; Eastern 0.90, Aegean 0.71, Central Anatolia 0.49, Thrace 0.37,
Mediterranean 0.34, Black Sea 0.31, South-Eastern 0.25 h), transcribed with human review of suspicious clips; 2.5 h
Google FLEURS tr train (CC-BY-4.0) so standard Turkish is not forgotten; 0.5 h generated noise/silence with empty
targets. Test recordings and speakers are excluded from training. Audio is not redistributed. Full settings are in the
"Teknik ayarlar" table above; one NVIDIA RTX 4090.
Licence
CC-BY-NC-4.0: free for research and personal use, no commercial use. You may use, adapt and share the model with attribution to this repository; use in a paid product or service needs permission. Released as a feasibility study. The base model Whisper is MIT; the standard Turkish data is Google FLEURS (CC-BY-4.0).
Limitations
Small test sets (5โ16 clips per region). No gain in Central Anatolia, a loss in the Mediterranean; Eastern is still hard. Occasional repetition loops on long or song-like audio. Test references went through the same labelling process as training labels, so part of the gain may be label-style match. Out of scope: Kurdish, Zaza, Arabic and other languages; speaker diarization; timestamps; high-stakes uses (medical, legal, identity).
About
Built by eqlx. How we fine-tune speech models on an institution's own audio and terms: small language models ยท contact center speech use case.
- Downloads last month
- 66
Model tree for eqlx-ai/whisper-large-v3-turbo-tr-aksan
Evaluation results
- regional WER (untuned base 32.5) on tr-stt-aksan regional holdout (92 clipsself-reported25.800
- WER (untuned base 5.3) on google/fleurs tr_tr test (300 clips)self-reported5.300



