added audio processing function

app.py

@@ -9,6 +9,18 @@ import requests
 import json
 import dotenv
 from transformers import AutoProcessor, SeamlessM4TModel
+import torchaudio
+dotenv.load_dotenv()
+
+
+AUDIO_SAMPLE_RATE = 16000.0
+MAX_INPUT_AUDIO_LENGTH = 60  # in seconds
+DEFAULT_TARGET_LANGUAGE = "English"
+
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+
+processor = AutoProcessor.from_pretrained("ylacombe/hf-seamless-m4t-large")
+model = SeamlessM4TModel.from_pretrained("ylacombe/hf-seamless-m4t-large").to(device)
 
 from lang_list import (
     LANGUAGE_NAME_TO_CODE,
@@ -17,61 +29,27 @@ from lang_list import (
     T2TT_TARGET_LANGUAGE_NAMES,
     TEXT_SOURCE_LANGUAGE_NAMES,
     LANG_TO_SPKR_ID,
-)
-
-dotenv.load_dotenv()
-
-DEFAULT_TARGET_LANGUAGE = "English"
-AUDIO_SAMPLE_RATE = 16000.0
-MAX_INPUT_AUDIO_LENGTH = 60  # in seconds
-
-
-def predict(
-    task_name: str,
-    audio_source: str,
-    input_audio_mic: str | None,
-    input_audio_file: str | None,
-    input_text: str | None,
-    source_language: str | None,
-    target_language: str,
-) -> tuple[tuple[int, np.ndarray] | None, str]:
-    task_name = task_name.split()[0]
-    source_language_code = LANGUAGE_NAME_TO_CODE[source_language] if source_language else None
-    target_language_code = LANGUAGE_NAME_TO_CODE[target_language]
-
-    if task_name in ["S2ST", "S2TT", "ASR"]:
-        if audio_source == "microphone":
-            input_data = input_audio_mic
-        else:
-            input_data = input_audio_file
-
-        arr, org_sr = torchaudio.load(input_data)
-        new_arr = torchaudio.functional.resample(arr, orig_freq=org_sr, new_freq=AUDIO_SAMPLE_RATE)
-        max_length = int(MAX_INPUT_AUDIO_LENGTH * AUDIO_SAMPLE_RATE)
-        if new_arr.shape[1] > max_length:
-            new_arr = new_arr[:, :max_length]
-            gr.Warning(f"Input audio is too long. Only the first {MAX_INPUT_AUDIO_LENGTH} seconds is used.")
-
-        
-        input_data = processor(audios = new_arr, sampling_rate=AUDIO_SAMPLE_RATE, return_tensors="pt").to(device)
-    else:
-        input_data = processor(text = input_text, src_lang=source_language_code, return_tensors="pt").to(device)
+)    
 
 
-    if task_name in ["S2TT", "T2TT"]:
-        tokens_ids = model.generate(**input_data, generate_speech=False, tgt_lang=target_language_code, num_beams=5, do_sample=True)[0].cpu().squeeze().detach().tolist()
-    else:
-        output = model.generate(**input_data, return_intermediate_token_ids=True, tgt_lang=target_language_code, num_beams=5, do_sample=True, spkr_id=LANG_TO_SPKR_ID[target_language_code][0])
-        
-        waveform = output.waveform.cpu().squeeze().detach().numpy()
-        tokens_ids = output.sequences.cpu().squeeze().detach().tolist()
-
+def process_speech(sound):
+    """
+    processing sound using seamless_m4t
+    """
+    # task_name = "T2TT"
+    arr, org_sr = torchaudio.load(sound)
+    target_language_code = LANGUAGE_NAME_TO_CODE[DEFAULT_TARGET_LANGUAGE]
+    new_arr = torchaudio.functional.resample(arr, orig_freq=org_sr, new_freq=AUDIO_SAMPLE_RATE)
+    max_length = int(MAX_INPUT_AUDIO_LENGTH * AUDIO_SAMPLE_RATE)
+    if new_arr.shape[1] > max_length:
+        new_arr = new_arr[:, :max_length]
+        gr.Warning(f"Input audio is too long. Only the first {MAX_INPUT_AUDIO_LENGTH} seconds is used.")
+    input_data = processor(audios = new_arr, sampling_rate=AUDIO_SAMPLE_RATE, return_tensors="pt").to(device)
+    tokens_ids = model.generate(**input_data, generate_speech=False, tgt_lang=target_language_code, num_beams=5, do_sample=True)[0].cpu().squeeze().detach().tolist()
     text_out = processor.decode(tokens_ids, skip_special_tokens=True)
 
-    if task_name in ["S2ST", "T2ST"]:
-        return (AUDIO_SAMPLE_RATE, waveform), text_out
-    else:
-        return None, text_out
+    return text_out
+    
 
 def convert_image_to_required_format(image):
     """