Upload 6 files

app.py

@@ -0,0 +1,11 @@
+from utils_trans import IndicTrans
+from utils_indic_seamless import INDIC_SEAMLESS
+from json import load as json_load
+from torch import device as Device
+from torch.cuda import is_available as cuda_is_available
+device = Device("cuda" if cuda_is_available() else "cpu")
+indictrans_model = IndicTrans(json_load(open("indictrans_conf.json")),"prajdabre/rotary-indictrans2-en-indic-dist-200M","prajdabre/rotary-indictrans2-indic-en-dist-200M","ai4bharat/indictrans2-indic-indic-dist-320M")
+indic_seamless_model = INDIC_SEAMLESS(json_load(open("indic_seamless_lang_conf.json")),"shethjenil/INDIC_SEAMLESS",device)
+
+import gradio as gr
+gr.TabbedInterface([gr.Interface(indictrans_model.translate,[gr.Textbox(label="Input Text"),gr.Dropdown(indictrans_model.all_lang, label="Source Language"),gr.Dropdown(indictrans_model.all_lang, label="Target Language"),],gr.Textbox(label="Result"),),gr.Interface(lambda files, lang: indic_seamless_model.speech2translate([i.name for i in files], lang),[gr.File(file_types=["audio"],label="Upload Audio Files",file_count="multiple",),gr.Dropdown(list(indic_seamless_model.lang_conf.keys()), label="Target Language"),],gr.List(label="Translations"),title="Audio Translation",),],["Indic Translation","Indic Audio Translation",],).launch()

indic_seamless_lang_conf.json

@@ -0,0 +1,15 @@
+{
+    "Assamese": "asm",
+    "Bengali": "ben",
+    "Gujarati": "guj",
+    "Hindi": "hin",
+    "Kannada": "kan",
+    "Malayalam": "mal",
+    "Marathi": "mar",
+    "Odia": "ory",
+    "Punjabi": "pan",
+    "Tamil": "tam",
+    "Telugu": "tel",
+    "Urdu": "urd",
+    "English": "eng"
+}

indictrans_conf.json

@@ -0,0 +1,36 @@
+[
+    "asm_Beng",
+    "awa_Deva",
+    "ben_Beng",
+    "bho_Deva",
+    "brx_Deva",
+    "doi_Deva",
+    "eng_Latn",
+    "gom_Deva",
+    "gon_Deva",
+    "guj_Gujr",
+    "hin_Deva",
+    "hne_Deva",
+    "kan_Knda",
+    "kas_Arab",
+    "kas_Deva",
+    "kha_Latn",
+    "lus_Latn",
+    "mag_Deva",
+    "mai_Deva",
+    "mal_Mlym",
+    "mar_Deva",
+    "mni_Beng",
+    "mni_Mtei",
+    "npi_Deva",
+    "ory_Orya",
+    "pan_Guru",
+    "san_Deva",
+    "sat_Olck",
+    "snd_Arab",
+    "snd_Deva",
+    "tam_Taml",
+    "tel_Telu",
+    "urd_Arab",
+    "unr_Deva"
+]

requirements.txt

@@ -0,0 +1,7 @@
+## for text translation
+git+https://github.com/VarunGumma/IndicTransToolkit
+
+## for speech translation
+transformers
+pydub
+numpy

utils_indic_seamless.py

@@ -0,0 +1,12 @@
+from transformers import SeamlessM4Tv2ForSpeechToText,SeamlessM4TTokenizer, SeamlessM4TFeatureExtractor
+from numpy import array as np_array,float32 as np_float32
+from pydub import AudioSegment
+
+class INDIC_SEAMLESS:
+    def __init__(self,lang_conf:dict[str,str],model,device):
+            self.seamless_model = SeamlessM4Tv2ForSpeechToText.from_pretrained(model).to(device)
+            self.seamless_processor = SeamlessM4TFeatureExtractor.from_pretrained(model)
+            self.seamless_tokenizer = SeamlessM4TTokenizer.from_pretrained(model)
+            self.lang_conf = lang_conf
+    def speech2translate(self,audio_paths, target_lang):
+        return self.seamless_tokenizer.batch_decode(self.seamless_model.generate(**self.seamless_processor([np_array(AudioSegment.from_file(path).set_channels(1).set_frame_rate(16000).get_array_of_samples(), dtype=np_float32) / 32768.0 for path in audio_paths], sampling_rate=16000, return_tensors="pt", padding=True).to("cpu"), tgt_lang=self.lang_conf[target_lang]), skip_special_tokens=True)

utils_trans.py

@@ -0,0 +1,27 @@
+import torch
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+from IndicTransToolkit.processor import IndicProcessor
+class IndicTrans:
+    def __init__(self,all_lang:list[str],en2indic,indic2en,indic2indic):
+        self.all_lang = all_lang
+        self.ip = IndicProcessor(inference=True)
+        self.indictrans_en2indic_tokenizer = AutoTokenizer.from_pretrained(en2indic, trust_remote_code=True)
+        self.indictrans_en2indic_model = AutoModelForSeq2SeqLM.from_pretrained(en2indic, trust_remote_code=True)
+        self.indictrans_indic2en_tokenizer = AutoTokenizer.from_pretrained(indic2en, trust_remote_code=True)
+        self.indictrans_indic2en_model = AutoModelForSeq2SeqLM.from_pretrained(indic2en, trust_remote_code=True)
+        self.indictrans_indic2indic_tokenizer = AutoTokenizer.from_pretrained(indic2indic, trust_remote_code=True)
+        self.indictrans_indic2indic_model = AutoModelForSeq2SeqLM.from_pretrained(indic2indic, trust_remote_code=True)
+    def _translate(self,model,tokenizer,input_list: list[str], source_lang: str, target_lang: str)->list[str]:
+        with torch.inference_mode():
+            outputs = model.generate(**tokenizer(self.ip.preprocess_batch(input_list, src_lang=source_lang, tgt_lang=target_lang, visualize=False),padding="longest",truncation=True,max_length=256,return_tensors="pt"), num_beams=5, num_return_sequences=1, max_length=256)
+        with tokenizer.as_target_tokenizer():
+            outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+        return self.ip.postprocess_batch(outputs, lang=target_lang)
+    def translate(self,input: str, source_lang: str, target_lang: str):
+        assert source_lang != target_lang and source_lang in self.all_lang and target_lang in self.all_lang
+        if source_lang == "eng_Latn":
+            return self._translate(self.indictrans_en2indic_model,self.indictrans_en2indic_tokenizer,[input],source_lang,target_lang)[0]
+        elif target_lang == "eng_Latn":
+            return self._translate(self.indictrans_indic2en_model,self.indictrans_indic2en_tokenizer,[input],source_lang,target_lang)[0]
+        else:
+            return self._translate(self.indictrans_indic2indic_model,self.indictrans_indic2indic_tokenizer,[input],source_lang,target_lang)[0]