Improve error messaging for ASR pipeline.

src/transformers/pipelines/automatic_speech_recognition.py

@@ -250,6 +250,10 @@ def preprocess(self, inputs, chunk_length_s=0, stride_length_s=None):
             raise ValueError("We expect a single channel audio input for AutomaticSpeechRecognitionPipeline")
 
         if chunk_length_s:
+            if self.type not in {"ctc", "ctc_with_lm"}:
+                raise ValueError(
+                    "`chunk_length_s` is only valid for CTC models, use other chunking options for other models"
+                )
             if stride_length_s is None:
                 stride_length_s = chunk_length_s / 6
 
@@ -264,10 +268,6 @@ def preprocess(self, inputs, chunk_length_s=0, stride_length_s=None):
             stride_left = int(round(stride_length_s[0] * self.feature_extractor.sampling_rate / align_to) * align_to)
             stride_right = int(round(stride_length_s[1] * self.feature_extractor.sampling_rate / align_to) * align_to)
 
-            if self.type not in {"ctc", "ctc_with_lm"}:
-                raise ValueError(
-                    "`chunk_length_s` is only valid for CTC models, use other chunking options for other models"
-                )
             if chunk_len < stride_left + stride_right:
                 raise ValueError("Chunk length must be superior to stride length")
 

tests/pipelines/test_pipelines_automatic_speech_recognition.py

@@ -118,9 +118,15 @@ def run_pipeline_test(self, speech_recognizer, examples):
                 },
             )
         else:
+            # Non CTC models cannot use chunk_length
+            with self.assertRaises(ValueError) as v:
+                outputs = speech_recognizer(audio, chunk_length_s=10)
+            self.assertEqual(v.exception, "")
+
             # Non CTC models cannot use return_timestamps
-            with self.assertRaises(ValueError):
+            with self.assertRaises(ValueError) as v:
                 outputs = speech_recognizer(audio, return_timestamps="char")
+            self.assertEqual(v.exception, "")
 
     @require_torch
     @slow
@@ -138,6 +144,17 @@ def test_small_model_pt(self):
         waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
         output = speech_recognizer(waveform)
         self.assertEqual(output, {"text": "(Applaudissements)"})
+        with self.assertRaises(ValueError) as v:
+            _ = speech_recognizer(waveform, chunk_length_s=10)
+        self.assertEqual(
+            str(v.exception),
+            "`chunk_length_s` is only valid for CTC models, use other chunking options for other models",
+        )
+
+        # Non CTC models cannot use return_timestamps
+        with self.assertRaises(ValueError) as v:
+            _ = speech_recognizer(waveform, return_timestamps="char")
+        self.assertEqual(str(v.exception), "We cannot return_timestamps yet on non-ctc models !")
 
     @require_torch
     def test_small_model_pt_seq2seq(self):