[CI] Compute speech WER directly with jiwer (#49773)
Chang Guo committed
Jul 25, 2026 at 17:51 UTC
7a6a5b3667160ab6c05fbda0503b12a1f3bff32e
1 file changed
+3
-5
tests/entrypoints/speech_to_text/correctness/test_transcription_api_correctness.py
+3
-5
@@ -17,7 +17,7 @@ import pytest
17
import soundfile
18
import torch
19
from datasets import Audio, load_dataset
20
-from evaluate import load
20
+from jiwer import wer
21
from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
22
23
from vllm.benchmarks.datasets.datasets import ASRDataset
@@ -202,8 +202,7 @@ def run_evaluation(
202
# Compute WER
203
predictions = [res[2] for res in results]
204
references = [res[3] for res in results]
205
- wer = load("wer")
206
- wer_score = 100 * wer.compute(references=references, predictions=predictions)
205
+ wer_score = 100 * wer(references, predictions)
206
print("WER:", wer_score)
207
return wer_score
208
@@ -302,8 +301,7 @@ def run_longform_evaluation(
301
302
predictions = [res[2] for res in results]
303
references = [res[3] for res in results]
305
- wer = load("wer")
306
- wer_score = 100 * wer.compute(references=references, predictions=predictions)
304
+ wer_score = 100 * wer(references, predictions)
305
print("WER:", wer_score)
306
return wer_score
307