Instructions to use Reza2kn/visualears-fastconformer-fa-full-ab-nvfp4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use Reza2kn/visualears-fastconformer-fa-full-ab-nvfp4 with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("Reza2kn/visualears-fastconformer-fa-full-ab-nvfp4") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
Add FP base comparison summary
Browse files
validation/nvfp4_vs_base_eval_summary.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n_eval": 200,
|
| 3 |
+
"base_wer": 0.1838006230529595,
|
| 4 |
+
"base_cer": 0.06584178498985802,
|
| 5 |
+
"base_wall_seconds": 6.194460868835449,
|
| 6 |
+
"base_peak_vram_mib": 588.24658203125,
|
| 7 |
+
"nvfp4_wer": 0.20327102803738317,
|
| 8 |
+
"nvfp4_cer": 0.0737525354969574,
|
| 9 |
+
"exact_normalized_transcript_match": 0.215,
|
| 10 |
+
"exact_matches": 43,
|
| 11 |
+
"word_position_agreement_rough": 0.7853986551392892,
|
| 12 |
+
"note": "Transcript-level comparison on same 200 FLEURS-fa clips; not CTC-logit argmax parity."
|
| 13 |
+
}
|