Add ThaiG2P v3 ONNX engine to transliterate

[Copilot]

ThaiG2P v3 is a char-level Transformer G2P model exported to ONNX, offering improved training data over v2 with a smaller model footprint.

What do these changes do

• pythainlp/corpus/thaig2p_v3_encoder.onnx — Bundled encoder ONNX model
• pythainlp/corpus/thaig2p_v3_decoder.onnx — Bundled decoder ONNX model
• pythainlp/corpus/thaig2p_v3_vocab.json — Bundled vocabulary (character-to-index mapping)
• pythainlp/corpus/default_db.json — Added entries for the three bundled model files
• pythainlp/transliterate/thaig2p_v3.py — New module: ONNX inference via onnxruntime with greedy decoder loop; loads model files from pythainlp/corpus via get_corpus_path()
• pythainlp/transliterate/core.py — Added thaig2p_v3 engine dispatch + docstring entry
• tests/extra/testx_transliterate.py — Added thaig2p_v3 test cases alongside existing v2/umt5 tests
• docs/api/transliterate.rst — Documented the new engine
• CHANGELOG.md — Logged under [Unreleased]

Usage

from pythainlp.transliterate import transliterate

transliterate("สวัสดี", engine="thaig2p_v3")
# '/sa˨˩.wat̚˨˩.diː˧/'

transliterate("ภาษาไทย", engine="thaig2p_v3")
# '/pʰaː˧.saː˩˩˦.tʰaj˧/'

What was wrong

ThaiG2P v3 existed upstream (wannaphong/thai-g2p-v3) but was not integrated into PyThaiNLP.

How this fixes it

Adds thaig2p_v3 as a first-class transliterate() engine. Unlike v2 (HuggingFace Transformers pipeline), v3 uses onnxruntime directly, matching the ONNX-first design of the upstream model. The three model files (thaig2p_v3_encoder.onnx, thaig2p_v3_decoder.onnx, thaig2p_v3_vocab.json) are bundled in pythainlp/corpus/ and registered in default_db.json, consistent with how thai2rom_onnx is packaged. The module loads them via get_corpus_path() with no custom download logic required.

Your checklist for this pull request

• Passed code styles and structures
• Passed code linting checks and unit test

[wannaphong]

@copilot Github v3: https://github.com/wannaphong/thai-g2p-v3

[wannaphong]

@copilot Example:

import onnxruntime as ort
import numpy as np
import json

# Load Vocab
with open("thai2ipa_vocab.json", "r", encoding="utf-8") as f:
    vocab = json.load(f)
char2idx = vocab["input_char2idx"]
idx2char = {int(k): v for k, v in vocab["target_idx2char"].items()}

# Load ONNX Sessions
encoder_sess = ort.InferenceSession("thai2ipa_encoder.onnx", providers=['CPUExecutionProvider'])
decoder_sess = ort.InferenceSession("thai2ipa_decoder.onnx", providers=['CPUExecutionProvider'])

def transliterate_onnx(text, max_len=50):
    # 1. Prepare Input
    src = [1] + [char2idx.get(c, 3) for c in text] + [2]  # <SOS>, chars, <EOS>
    src_tensor = np.array([src], dtype=np.int64)
    
    # 2. Run Encoder
    enc_outputs = encoder_sess.run(None, {"src": src_tensor})
    memory, src_pad_mask = enc_outputs[0], enc_outputs[1]
    
    # 3. Autoregressive Decoder Loop (Greedy Search)
    trg_indexes = [1] # Start with <SOS>
    
    for _ in range(max_len):
        trg_tensor = np.array([trg_indexes], dtype=np.int64)
        
        # Run Decoder
        dec_outputs = decoder_sess.run(None, {
            "trg": trg_tensor,
            "memory": memory,
            "src_pad_mask": src_pad_mask
        })
        
        # dec_outputs[0] shape: (Batch, Trg_Len, Vocab_Size)
        # Get the highest probability token for the LAST generated step
        next_token_logits = dec_outputs[0][0, -1, :]
        next_token = int(np.argmax(next_token_logits))
        
        if next_token == 2: # <EOS> reached
            break
            
        trg_indexes.append(next_token)
        
    # 4. Decode to String
    result = "".join([idx2char[idx] for idx in trg_indexes[1:]])
    return result

print(transliterate_onnx("สวัสดี"))
print(transliterate_onnx("สวัสดีค่ะ"))
print(transliterate_onnx("คน"))
print(transliterate_onnx("ดี"))
print(transliterate_onnx("วรรณพงษ์"))
print(transliterate_onnx("ราชกิจ"))

[wannaphong]

@copilot Include model files to pythainlp/corpus

[Copilot]

@copilot Include model files to pythainlp/corpus

Done in commit 7749340. The three model files (thaig2p_v3_encoder.onnx, thaig2p_v3_decoder.onnx, thaig2p_v3_vocab.json) are now bundled in pythainlp/corpus/ with entries added to default_db.json. The thaig2p_v3.py module now uses get_corpus_path() to load them, consistent with how thai2rom_onnx works.

[coveralls]

coverage: 66.26% (-0.4%) from 66.633%
when pulling a21b0f9 on copilot/add-thai-g2p-v3-to-pythainlp
into 3152a36 on dev.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud