improve

avg_wer.py

@@ -0,0 +1,24 @@
+#!/usr/bin/env python3
+import sys
+
+file = sys.argv[1]
+key = sys.argv[2]
+
+with open(file, "r") as f:
+    lines = f.readlines()
+
+wers = []
+count = 0
+
+for line in lines:
+    if key in line:
+        wers.append(float(line.split(key)[-1].lstrip().rstrip()))
+        count += 1
+
+mean = sum(wers) / len(wers)
+var = sum([(w - mean)**2 for w in wers]) / len(wers)
+
+
+print("mean", mean)
+print("var", var)
+

hf_whisper_meanwhile.py

@@ -42,7 +42,7 @@ for audio, label in zip(audios, labels):
     if inputs["input_features"].shape[-1] < 3000:
         continue
 
-    # result = model_orig.transcribe(audio.astype(dtype=np.float32), condition_on_previous_text=DO_COND, temperature=0.0, logprob_threshold=None, compression_ratio_threshold=None, no_speech_threshold=None)
+    result = model_orig.transcribe(audio.astype(dtype=np.float32), condition_on_previous_text=DO_COND, temperature=0.0, logprob_threshold=None, compression_ratio_threshold=None, no_speech_threshold=None)
 
     gen_length = 448
     result_hf = model.generate(**inputs, condition_on_prev_tokens=DO_COND, max_new_tokens=gen_length, return_timestamps=True)