Automatic Speech Recognition
Transformers
Safetensors
Hungarian
whisper
Generated from Trainer
hungarian
Eval Results (legacy)
Instructions to use sarpba/whisper-base-hungarian_v3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use sarpba/whisper-base-hungarian_v3 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="sarpba/whisper-base-hungarian_v3")# Load model directly from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("sarpba/whisper-base-hungarian_v3") model = AutoModelForSpeechSeq2Seq.from_pretrained("sarpba/whisper-base-hungarian_v3", device_map="auto") - Notebooks
- Google Colab
- Kaggle
README: teljes összehasonlító eredmények (FLEURS + CV17 + zajos, beam 1-5)
Browse files
README.md
CHANGED
|
@@ -13,7 +13,7 @@ datasets:
|
|
| 13 |
metrics:
|
| 14 |
- wer
|
| 15 |
model-index:
|
| 16 |
-
- name: Whisper Base Hungarian
|
| 17 |
results:
|
| 18 |
- task:
|
| 19 |
name: Automatic Speech Recognition
|
|
@@ -30,29 +30,41 @@ model-index:
|
|
| 30 |
value: 29.81
|
| 31 |
---
|
| 32 |
|
| 33 |
-
# Whisper Base Hungarian
|
| 34 |
|
| 35 |
Magyar nyelvre finomhangolt [openai/whisper-base](https://huggingface.co/openai/whisper-base) modell, zaj-augmentált tréninggel.
|
| 36 |
|
| 37 |
-
A
|
| 38 |
|
| 39 |
## Eredmények
|
| 40 |
|
| 41 |
### Tiszta anyag — FLEURS hu_hu test (normalizált WER)
|
| 42 |
|
| 43 |
-
| Modell | beam=1 | beam=5 |
|
| 44 |
-
|---|---|---|
|
| 45 |
-
|
|
| 46 |
-
| whisper-base-
|
|
|
|
| 47 |
|
| 48 |
-
###
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
| Modell | WER |
|
| 51 |
|---|---|
|
| 52 |
-
|
|
|
|
|
| 53 |
| whisper-base-hungarian_v1 | 56.2 |
|
| 54 |
|
| 55 |
-
A
|
| 56 |
|
| 57 |
## Model description
|
| 58 |
|
|
@@ -117,7 +129,7 @@ from transformers import pipeline
|
|
| 117 |
|
| 118 |
pipe = pipeline(
|
| 119 |
"automatic-speech-recognition",
|
| 120 |
-
model="sarpba/whisper-base-
|
| 121 |
)
|
| 122 |
result = pipe("magyar_beszed.wav", generate_kwargs={"language": "hungarian", "task": "transcribe"})
|
| 123 |
print(result["text"])
|
|
|
|
| 13 |
metrics:
|
| 14 |
- wer
|
| 15 |
model-index:
|
| 16 |
+
- name: Whisper Base Hungarian v3
|
| 17 |
results:
|
| 18 |
- task:
|
| 19 |
name: Automatic Speech Recognition
|
|
|
|
| 30 |
value: 29.81
|
| 31 |
---
|
| 32 |
|
| 33 |
+
# Whisper Base Hungarian v3
|
| 34 |
|
| 35 |
Magyar nyelvre finomhangolt [openai/whisper-base](https://huggingface.co/openai/whisper-base) modell, zaj-augmentált tréninggel.
|
| 36 |
|
| 37 |
+
A v3 a v1 (whisper-base-hungarian_v1) továbbfejlesztett változata: a tréning során a tiszta hanganyag mellett véletlenszerűen zajjal kevert mintákon is tanult, így a **zajos környezetben lényegesen robusztusabb** a v1-nél, miközben tiszta anyagon is tartja (vagy éppen meg is haladja) a v1 teljesítményét.
|
| 38 |
|
| 39 |
## Eredmények
|
| 40 |
|
| 41 |
### Tiszta anyag — FLEURS hu_hu test (normalizált WER)
|
| 42 |
|
| 43 |
+
| Modell | beam=1 | beam=2 | beam=3 | beam=4 | beam=5 |
|
| 44 |
+
|---|---|---|---|---|---|
|
| 45 |
+
| whisper-hu-base-finetuned-V2 (régi) | **25.79** | **24.26** | **23.80** | **23.77** | **23.67** |
|
| 46 |
+
| **whisper-base-hungarian_v3 (ez)** | 29.81 | 29.07 | 28.84 | 28.57 | 29.28 |
|
| 47 |
+
| whisper-base-hungarian_v1 | 30.91 | 29.19 | 28.94 | 29.04 | 28.77 |
|
| 48 |
|
| 49 |
+
### Common Voice 17.0 hu test (normalizált WER)
|
| 50 |
+
|
| 51 |
+
| Modell | beam=1 | beam=2 | beam=3 | beam=4 | beam=5 |
|
| 52 |
+
|---|---|---|---|---|---|
|
| 53 |
+
| whisper-hu-base-finetuned-V2 (régi) | **14.73** | **14.03** | **13.90** | **13.85** | **13.81** |
|
| 54 |
+
| **whisper-base-hungarian_v3 (ez)** | 21.80 | 21.23 | 21.11 | 21.06 | 21.09 |
|
| 55 |
+
| whisper-base-hungarian_v1 | 21.48 | 20.75 | 20.57 | 20.54 | 20.51 |
|
| 56 |
+
|
| 57 |
+
_Megjegyzés: a Common Voice 17.0 train+valid spliteket a régi V2 tréningje tartalmazta, ezért ott az eredményei nem teljesen függetlenek._
|
| 58 |
+
|
| 59 |
+
### Zajos anyag (~10 dB MUSAN zaj, 905 minta, beam=1)
|
| 60 |
|
| 61 |
| Modell | WER |
|
| 62 |
|---|---|
|
| 63 |
+
| whisper-hu-base-finetuned-V2 (régi) | **41.6** |
|
| 64 |
+
| **whisper-base-hungarian_v3 (ez)** | 51.0 |
|
| 65 |
| whisper-base-hungarian_v1 | 56.2 |
|
| 66 |
|
| 67 |
+
A v3 a v1-hez képest minden készleten jobb (FLEURS: −1.1…−0.3, CV17: −0.3…−0.7, zajos: **−5.2 pont**). A régi V2 (2000 órás, CV17-tel bővített adat, normalizált célok) minden készleten jobb nála — elsősorban a lényegesen nagyobb és változatosabb tréningadata miatt.
|
| 68 |
|
| 69 |
## Model description
|
| 70 |
|
|
|
|
| 129 |
|
| 130 |
pipe = pipeline(
|
| 131 |
"automatic-speech-recognition",
|
| 132 |
+
model="sarpba/whisper-base-hungarian_v3",
|
| 133 |
)
|
| 134 |
result = pipe("magyar_beszed.wav", generate_kwargs={"language": "hungarian", "task": "transcribe"})
|
| 135 |
print(result["text"])
|