Prep for Voice Steering feature (#141)

* Prep for Voice Steering feature

Credits:

1. ylacombe
- Add input_values to DACModel
- dac_wrapper/modeling_dac.py
- #110 (comment)

2. stg2015
- Delay mask adjustment for input_values
- modeling_parler_tts.py
- #81 (comment)

* Prep for voice steering/cloning w/ fix for non-streaming generation

* Applied simpler input handling per Guppy16's suggestion

* Applied Guppy16's suggested optimization

* Applied Guppy17's suggested optimization for voice steering

* Update parler_tts/modeling_parler_tts.py

---------

Co-authored-by: apresence <[email protected]>
Co-authored-by: Yoach Lacombe <[email protected]>

Author: 3 people

parler_tts/dac_wrapper/modeling_dac.py

@@ -12,6 +12,9 @@
 class DACModel(PreTrainedModel):
     config_class = DACConfig
 
+    # Set main input to 'input_values' for voice steering
+    main_input_name = "input_values"
+
     def __init__(self, config):
         super().__init__(config)
 

parler_tts/modeling_parler_tts.py

@@ -3483,13 +3483,17 @@ def generate(
         # Apply the pattern mask to the final ids
         output_ids = self.decoder.apply_delay_pattern_mask(output_ids, model_kwargs["decoder_delay_pattern_mask"])
 
-        # Revert the pattern delay mask by filtering the eos and bos token ids from the delay pattern mask
-        _, mask = self.decoder.build_delay_pattern_mask(
-            input_ids,
-            bos_token_id=generation_config._bos_token_tensor,
-            pad_token_id=generation_config._pad_token_tensor,
-            max_length=output_ids.shape[1],
-        )
+        if "input_values" in model_kwargs:
+            # Handle input_values for voice steering
+            mask = output_ids
+        else:
+            # Revert the pattern delay mask by filtering the eos and bos token ids from the delay pattern mask
+            _, mask = self.decoder.build_delay_pattern_mask(
+                input_ids,
+                bos_token_id=generation_config.bos_token_id,
+                pad_token_id=generation_config.pad_token_id,
+                max_length=output_ids.shape[1],
+            )
 
         mask = (mask != generation_config.bos_token_id) & (mask != generation_config.pad_token_id)
         output_ids = output_ids[mask].reshape(batch_size, self.decoder.num_codebooks, -1)