asr model approach

app.py

@@ -5,6 +5,7 @@ import numpy as np
 import torch
 
 from transformers import VitsModel, VitsTokenizer, pipeline
+from transformers import WhisperForConditionalGeneration, WhisperProcessor
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
@@ -12,16 +13,35 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
 target_language = "fr"
 
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+# asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+whisper_model_name = "openai/whisper-base"
+whisper_processor = WhisperProcessor.from_pretrained(whisper_model_name)
+whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_name)
+decoder_ids = whisper_processor.get_decoder_prompt_ids(language=target_language, task="transcribe")
 
 # load text-to-speech checkpoint
 model = VitsModel.from_pretrained("facebook/mms-tts-fra")
 tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-fra")
 
 
+# def translate(audio):
+#     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": target_language})
+#     return outputs["text"]
+
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": target_language})
-    return outputs["text"]
+    if isinstance(audio, str):
+        # Account for recorded audio
+        audio = {
+            "path": audio,
+            "sampling_rate": 16_000,
+            "array": librosa.load(audio, sr=16_000)[0]
+        }
+    elif audio["sampling_rate"] != 16_000:
+        audio["array"] = librosa.resample(audio["array"], audio["sampling_rate"], 16_000)
+    input_features = whisper_processor(audio["array"], sampling_rate=16000, return_tensors="pt").input_features
+    predicted_ids = whisper_model.generate(input_features, forced_decoder_ids=decoder_ids)
+    translated_text = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+    return translated_text
 
 
 def synthesise(text):