feat: add optional G2P conversion using formog2p

Add a checkbox to toggle G2P conversion (default off). When enabled,
formog2p converts Hakka Chinese characters to uppercase pinyin before
passing to the model, with dialect-aware lang_group mapping.

app.py

@@ -24,6 +24,14 @@ DIALECT_LABELS = [
     "客語詔安腔",
     "客語南四縣腔",
 ]
+DIALECT_TO_LANG_GROUP = {
+    "客語四縣腔": "hak_sx",
+    "客語海陸腔": "hak_hl",
+    "客語大埔腔": "hak_dp",
+    "客語饒平腔": "hak_rp",
+    "客語詔安腔": "hak_za",
+    "客語南四縣腔": "hak_nsx",
+}
 DEFAULT_SPEED = 1.0
 DEFAULT_STEPS = 32
 EXAMPLES = [
@@ -34,6 +42,7 @@ EXAMPLES = [
         "恁早。",
         DEFAULT_SPEED,
         DEFAULT_STEPS,
+        False,
     ],
     [
         "客語四縣腔",
@@ -42,6 +51,7 @@ EXAMPLES = [
         "你今晡日著到恁派頭。",
         DEFAULT_SPEED,
         DEFAULT_STEPS,
+        False,
     ],
     [
         "客語四縣腔",
@@ -50,6 +60,7 @@ EXAMPLES = [
         "你今晡日著到恁派頭。",
         DEFAULT_SPEED,
         DEFAULT_STEPS,
+        False,
     ],
 ]
 
@@ -142,6 +153,14 @@ def startup_status() -> str:
     )
 
 
+def apply_g2p(text: str, dialect: str) -> str:
+    from formog2p.hakka.g2p import g2p
+
+    lang_group = DIALECT_TO_LANG_GROUP.get(dialect, "hak_sx")
+    result = g2p(text, lang_group=lang_group, pronunciation_type="pinyin")
+    return " ".join(result.pronunciations).upper()
+
+
 def validate_inputs(
     dialect: str | None,
     text: str,
@@ -175,6 +194,7 @@ def synthesize(
     ref_text: str,
     speed: float,
     num_step: int,
+    use_g2p: bool,
 ) -> tuple[tuple[int, np.ndarray] | None, str]:
     error = validate_inputs(dialect, text, ref_audio, ref_text)
     if error:
@@ -188,6 +208,12 @@ def synthesize(
         return None, startup_status()
 
     try:
+        input_text = text.strip()
+        g2p_note = ""
+        if use_g2p:
+            input_text = apply_g2p(input_text, dialect)
+            g2p_note = f"；G2P 轉換：{input_text}"
+
         generation_config = RUNTIME.generation_config_cls(
             num_step=int(num_step),
             guidance_scale=2.0,
@@ -201,7 +227,7 @@ def synthesize(
             preprocess_prompt=True,
         )
         generate_kwargs: dict[str, Any] = {
-            "text": text.strip(),
+            "text": input_text,
             "voice_clone_prompt": voice_clone_prompt,
             "instruct": dialect,
             "generation_config": generation_config,
@@ -216,7 +242,7 @@ def synthesize(
 
         return (
             to_audio_output(audio[0], int(RUNTIME.sampling_rate or 24000)),
-            f"合成完成。腔調：{dialect}；speed={speed:.2f}；steps={int(num_step)}",
+            f"合成完成。腔調：{dialect}；speed={speed:.2f}；steps={int(num_step)}{g2p_note}",
         )
     except Exception as exc:
         return None, f"合成失敗：{type(exc).__name__}: {exc}"
@@ -260,6 +286,11 @@ def build_demo() -> gr.Blocks:
                         lines=2,
                         placeholder="請填寫參考音檔對應的逐字文本。",
                     )
+                    use_g2p = gr.Checkbox(
+                        value=False,
+                        label="使用 G2P 轉換",
+                        info="勾選後會先用 formog2p 將漢字轉成拼音（大寫）再輸入模型；不勾選則直接輸入原文。",
+                    )
                     with gr.Accordion("進階設定", open=False):
                         speed = gr.Slider(
                             minimum=0.5,
@@ -293,13 +324,13 @@ def build_demo() -> gr.Blocks:
 
             submit.click(
                 fn=synthesize,
-                inputs=[dialect, text, ref_audio, ref_text, speed, num_step],
+                inputs=[dialect, text, ref_audio, ref_text, speed, num_step, use_g2p],
                 outputs=[output_audio, status],
             )
 
             gr.Examples(
                 examples=EXAMPLES,
-                inputs=[dialect, text, ref_audio, ref_text, speed, num_step],
+                inputs=[dialect, text, ref_audio, ref_text, speed, num_step, use_g2p],
                 label="範例",
             )
 

requirements.txt

@@ -1,3 +1,4 @@
 git+https://github.com/txya900619/OmniVoice-hakka.git
 torch==2.8.0
-torchaudio==2.8.0
+torchaudio==2.8.0
+formog2p