Update README.md

README.md

@@ -33,61 +33,150 @@ from huggingface_hub import login
 login("TOKEN")
 ```
 
-**Load base model and BRDialect**<br>
+**Load BanglaIPA model**<br>
 ```python
-## BRDialect
-from huggingface_hub import hf_hub_download
+## BanglaIPA
+from huggingface_hub import snapshot_download
+import os
 
-kenlm_model_path = hf_hub_download(repo_id="Jakir057/BRDialect", filename="BRDialect/5gram_kenlm.arpa")
-state_dict_path = hf_hub_download(repo_id="Jakir057/BRDialect", filename="BRDialect/wav2vec2_bangla_regional_dialect.pth")
-```
-```python
-from transformers import AutoProcessor, AutoModelForCTC, Wav2Vec2ProcessorWithLM
-import torch
-import numpy as np
-import pyctcdecode
-import librosa
-
-base_model_id = "ai4bharat/indicwav2vec_v1_bengali"
-processor = AutoProcessor.from_pretrained(base_model_id)
-model = AutoModelForCTC.from_pretrained(base_model_id)
-model.load_state_dict(torch.load(state_dict_path)["model"])
-
-vocab_dict = processor.tokenizer.get_vocab()
-sorted_vocab_dict = {k: v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])}
-decoder = pyctcdecode.build_ctcdecoder(
-list(sorted_vocab_dict.keys()),
-str(kenlm_model_path)
+local_dir = snapshot_download(
+    repo_id="Jakir057/BanglaIPA"
 )
-processor_with_lm = Wav2Vec2ProcessorWithLM(
-feature_extractor=processor.feature_extractor,
-tokenizer=processor.tokenizer,
-decoder=decoder
-)
-model.freeze_feature_encoder()
-model.eval()
+print(local_dir)
+
+MODEL_PATH = os.path.join(local_dir, "BanglaIPA")
+print(f"Model path={MODEL_PATH}")
 ```
 
 ## Transcription Generation
 ```python
-sampling_rate = 16000
-path = "AUDIO_PATH"
-frame, sr = librosa.load(path, sr=sampling_rate, mono=True)
-
-inputs = processor(
-frame,
-sampling_rate=sampling_rate,
-return_tensors="pt",
-padding=False
-)
-
-with torch.no_grad():
-logits = model(inputs.input_values.to("cpu")).logits
-
-np_logits = logits.squeeze(0).cpu().numpy()
-result = processor_with_lm.decode(np_logits, beam_width=256)
-text = result.text
-print(f"Transcription={text}")
+import tensorflow as tf
+from tensorflow.keras.layers import TextVectorization
+import numpy as np
+import os
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
+
+def get_vocab():
+  """
+  Returns sorted list of Bengali characters, IPA characters, special tokens and other characters seen in the training set.
+  """
+  vb = ['', '[UNK]', '[start]', '[end]', 'া', 'র', '্', 'ে', 'ি', 'ন', 'ক', 'ব', 'স', 'ল', 'ত', 'ম', 'প', 'ু', 'দ', 'ট', 'য়', 'জ', '।', 'ো', 'গ', 'হ', 'য', 'শ', 'ী', 'ই', 'চ', 'ভ', 'আ', 'ও', 'ছ', 'ষ', 'ড', 'ফ', 'অ', 'ধ', 'খ', 'ড়', 'উ', 'ণ', 'এ', 'থ', 'ং', 'ঁ', 'ূ', 'ৃ', 'ঠ', 'ঘ', 'ঞ', 'ঙ', 'ৌ', '‘', 'ৎ', 'ঝ', 'ৈ', '়', 'ঢ', 'ঃ', 'ঈ', '\u200c', 'ৗ', 'a', 'ঐ', 'd', 'w', 'ঋ', 'i', 'e', 't', 's', 'n', 'm', 'b', '“', 'u', 'r', 'œ', 'o', '–', 'ঊ', 'ঢ়', 'Í', 'g', 'p', '\xad', 'h', 'c', 'l', 'ঔ', 'ƒ', '”', 'Ñ', '¡', 'y', 'j', 'f', '→', '—', 'ø', 'è', '¦', '¥', 'x', 'v', 'k']
+  vipa = ['', '[UNK]', '[start]', '[end]', 'ɐ', 'ɾ', 'i', 'o', 'e', '̪', 't', 'n', 'k', 'ɔ', 'ʃ', 'b', 'd', 'l', 'u', 'p', 'm', 'ʰ', 'ɟ', '͡', '̯', 'g', 'ʱ', '।', 'c', 'ʲ', 'h', 's', 'ŋ', 'ɛ', 'ɽ', '̃', 'ʷ', '‘', '“', '–', '”', '—', 'w', 'j']
+  v = vb + vipa
+  s = set()
+  for ch in v:
+    s.add(ch)
+  vocab = sorted(list(s))
+  return vocab
+
+def get_vectorization():
+  """
+  Performs vectorization.
+  """
+  vocab = get_vocab()
+  vocab_size = len(vocab)
+  sequence_length = 64
+  bn_vectorization = TextVectorization(
+      max_tokens=vocab_size, output_mode="int", output_sequence_length=sequence_length,
+      vocabulary=vocab
+  )
+  ipa_vectorization = TextVectorization(
+      max_tokens=vocab_size,
+      output_mode="int",
+      output_sequence_length=sequence_length + 1,
+      vocabulary=vocab
+  )
+  return bn_vectorization, ipa_vectorization
+
+def decode_sequence(input_sentence, bn_vectorization, ipa_vectorization, banglaipa_model):
+    """
+    Generate IPA for subword.
+    
+    Args:
+      - input_sentence (str): Synthetic sentence where every adjacent characters has a space between them.
+      - bn_vectorization: TextVectorization
+      - en_vectorization: TextVectorization
+      - banglaipa_model: Transformer model
+    Returns:
+      - str: String of IPA characters and special tokens where adjacent characters are separated with a space.
+    """
+    max_decoded_sentence_length = 64
+    spa_vocab = ipa_vectorization.get_vocabulary()
+    spa_index_lookup = dict(zip(range(len(spa_vocab)), spa_vocab))
+    tokenized_input_sentence = bn_vectorization([input_sentence])
+    decoded_sentence = '[start]'
+    for i in range(max_decoded_sentence_length):
+        tokenized_target_sentence = ipa_vectorization([decoded_sentence])[:, :-1]
+        predictions = banglaipa_model([tokenized_input_sentence, tokenized_target_sentence])
+        sampled_token_index = np.argmax(predictions[0, i, :])
+        sampled_token = spa_index_lookup[sampled_token_index]
+        decoded_sentence += " " + sampled_token
+        if sampled_token == '[UNK]':
+            break
+    return decoded_sentence
+
+def sentence_to_word(sentence):
+  """
+  Generate word from synthetic sentence by removing spaces between adjacent characters.
+
+  Args:
+    - sentence (str): Synthetic sentence.
+  Returns:
+    - str: subword/word
+  """
+  trg=''
+  for ch in sentence:
+      if ch != " ":
+        trg += ch
+  return trg
+
+def word_to_sentence(word):
+  """
+  Generate synthetic sentence from word by inserting spaces between adjacent characters.
+
+  Args:
+    - word (str): subword/word segement
+  Returns:
+    - str: Synthetic sentence
+  """
+  sentence = ""
+  for ch in word:
+    sentence += (ch + " ")
+  return sentence
+
+def get_subword2ipa(word, bn_vectorization, ipa_vectorization, banglaipa_model):
+    translated = decode_sequence(word_to_sentence(word), bn_vectorization, ipa_vectorization, banglaipa_model)
+    trg = sentence_to_word(translated)
+    trg = trg[7:]
+    trg = trg[:-5]
+    return trg
+
+if __name__ == "__main__":
+    path = MODEL_PATH
+    banglaipa_model=tf.saved_model.load(path)
+    print("BanglaIPA model loaded.")
+    bn_vectorization, ipa_vectorization = get_vectorization()
+    text = "একটি বাছাই করুন গণিত প্রথম গণিত দ্বিতীয় পত্র"
+    ipa = ""
+    words = text.split(" ")
+    for word in words:
+        trg = get_subword2ipa(word, bn_vectorization, ipa_vectorization, banglaipa_model)
+        print(word, trg)
+        ipa += (trg + " ")
+    print(f"IPA={ipa}")
+
+## python inference.py
+# # Output:
+# BanglaIPA model loaded.
+# একটি ekti
+# বাছাই bɐcʰɐ͡i̯
+# করুন koɾun
+# গণিত gonit̪o
+# প্রথম pɾot̪ʰom
+# গণিত gonit̪o
+# দ্বিতীয় d̪it̪iʲo
+# পত্র pɔt̪ɾo
+# IPA=ekti bɐcʰɐ͡i̯ koɾun gonit̪o pɾot̪ʰom gonit̪o d̪it̪iʲo pɔt̪ɾo
 ```
 
 ## Citation
@@ -95,12 +184,12 @@ print(f"Transcription={text}")
 ```
 @misc
 {hasan2026banglaiparobusttexttoipatranscription,
-title={BanglaIPA: Towards Robust Text-to-IPA Transcription with Contextual Rewriting in Bengali},
-author={Jakir Hasan and Shrestha Datta and Md Saiful Islam and Shubhashis Roy Dipta and Ameya Debnath},
-year={2026},
-eprint={2601.01778},
-archivePrefix={arXiv},
-primaryClass={cs.CL},
-url={https://arxiv.org/abs/2601.01778},
+  title={BanglaIPA: Towards Robust Text-to-IPA Transcription with Contextual Rewriting in Bengali},
+  author={Jakir Hasan and Shrestha Datta and Md Saiful Islam and Shubhashis Roy Dipta and Ameya Debnath},
+  year={2026},
+  eprint={2601.01778},
+  archivePrefix={arXiv},
+  primaryClass={cs.CL},
+  url={https://arxiv.org/abs/2601.01778},
 }
 ```