Add inputs_to_logits_ratio to LasrCTCConfig (#42720)

* Add inputs_to_logits_ratio to LasrCTCConfig

* changes

* nit

* update

* Add an _align_to property to unify stride computation in AutomaticSpeechRecognition pipeline

---------

Co-authored-by: Eustache Le Bihan <eulebihan@gmail.com>
Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com>

Author: kho

src/transformers/models/lasr/configuration_lasr.py

@@ -240,5 +240,9 @@ def from_encoder_config(cls, encoder_config: LasrEncoderConfig, **kwargs):
 
         return cls(encoder_config=encoder_config.to_dict(), **kwargs)
 
+    @property
+    def inputs_to_logits_ratio(self):
+        return self.encoder_config.subsampling_conv_stride**2
+
 
 __all__ = ["LasrEncoderConfig", "LasrCTCConfig"]

src/transformers/models/lasr/modular_lasr.py

@@ -291,6 +291,10 @@ def __init__(
             **kwargs,
         )
 
+    @property
+    def inputs_to_logits_ratio(self):
+        return self.encoder_config.subsampling_conv_stride**2
+
 
 class LasrEncoderSubsampling(nn.Module):
     def __init__(self, config: LasrEncoderConfig):

src/transformers/pipelines/automatic_speech_recognition.py

@@ -350,6 +350,20 @@ def _sanitize_parameters(
 
         return preprocess_params, forward_params, postprocess_params
 
+    @property
+    def _align_to(self):
+        """Sample stride per output."""
+        # XXX: Carefully, this variable will not exist in `seq2seq` setting.
+        # Currently chunking is not possible at this level for `seq2seq` so
+        # it's ok.
+        align_to = getattr(self.model.config, "inputs_to_logits_ratio", 1)
+        if self.model.config.model_type == "lasr_ctc":
+            # TODO: find a standard for that but not easy because input length -> mel length depends on the feature extractor
+            # specific way of doing it
+            # means the model take mel features as input, we align according to the hop length
+            align_to *= self.feature_extractor.hop_length
+        return align_to
+
     def preprocess(self, inputs, chunk_length_s=0, stride_length_s=None):
         if isinstance(inputs, str):
             if inputs.startswith("http://") or inputs.startswith("https://"):
@@ -444,10 +458,7 @@ def preprocess(self, inputs, chunk_length_s=0, stride_length_s=None):
             if isinstance(stride_length_s, (int, float)):
                 stride_length_s = [stride_length_s, stride_length_s]
 
-            # XXX: Carefully, this variable will not exist in `seq2seq` setting.
-            # Currently chunking is not possible at this level for `seq2seq` so
-            # it's ok.
-            align_to = getattr(self.model.config, "inputs_to_logits_ratio", 1)
+            align_to = self._align_to
             chunk_len = int(round(chunk_length_s * self.feature_extractor.sampling_rate / align_to) * align_to)
             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)
@@ -567,7 +578,7 @@ def _forward(self, model_inputs, return_timestamps=False, **generate_kwargs):
                 # Send stride to `postprocess`.
                 # it needs to be handled there where
                 # the pieces are to be concatenated.
-                ratio = 1 / self.model.config.inputs_to_logits_ratio
+                ratio = 1 / self._align_to
                 if isinstance(stride, tuple):
                     out["stride"] = rescale_stride([stride], ratio)[0]
                 else:
@@ -650,11 +661,12 @@ def postprocess(
 
         if return_timestamps and self.type not in {"seq2seq", "seq2seq_whisper"}:
             chunks = []
+            align_to = self._align_to
             for item in offsets:
-                start = item["start_offset"] * self.model.config.inputs_to_logits_ratio
+                start = item["start_offset"] * align_to
                 start /= self.feature_extractor.sampling_rate
 
-                stop = item["end_offset"] * self.model.config.inputs_to_logits_ratio
+                stop = item["end_offset"] * align_to
                 stop /= self.feature_extractor.sampling_rate
 
                 chunks.append({"text": item[return_timestamps], "timestamp": (start, stop)})