Automatically handle long text in both voice modes

README.md

@@ -20,9 +20,10 @@ models:
 
 ## 模式
 
-- **內建語者**：預設為內建語者 A；Speaker B 使用模型庫 `checkpoints/speaker_b_embedding.pt`，固定 seed 56789 與 projector scale 1.125。
-- **參考音色**：錄音直接送入模型原生的 raw reference-audio conditioning，不經 ECAPA centroid；固定 seed 1337，每個文字分段仍只生成一次。
-- **長文**：以最多 80 個字元快速分段，各分段單次生成後加入短靜音串接。
+- **內建語者**：預設為內建語者 A；Speaker B 使用模型庫 `checkpoints/speaker_b_embedding.pt`，固定 seed 56789 與 projector scale 1.125。長文會自動以最多 80 個字元分段。
+- **參考音色**：錄音直接送入模型原生的 raw reference-audio conditioning，不經 ECAPA centroid；固定 seed 1337。長文同樣自動分段。
+
+兩種模式的每個文字分段都只生成一次，分段間加入短靜音串接；不需要切換獨立長文模式。
 
 TTS model snapshot 固定為
 `6f7cab914a1e27c56b504ec663c0144dc25cc0a3`，預設 CFG 為 3.0，NFE 為 10。

app.py

@@ -24,18 +24,39 @@ os.environ["MAMBA_DETERMINISTIC"] = "0"
 
 INTERACTIVE_GPU_LEASE_SECONDS = 60
 LONGFORM_GPU_LEASE_SECONDS = 120
+AUTO_LONG_TEXT_CHARS = 80
+
+
+def _gpu_lease_seconds(text: object) -> int:
+    """Reserve the long-form lease only when automatic chunking needs it."""
+
+    return (
+        LONGFORM_GPU_LEASE_SECONDS
+        if len(str(text)) > AUTO_LONG_TEXT_CHARS
+        else INTERACTIVE_GPU_LEASE_SECONDS
+    )
+
+
+def _speaker_gpu_duration(text, speaker, cfg, steps, speed) -> int:
+    del speaker, cfg, steps, speed
+    return _gpu_lease_seconds(text)
+
+
+def _reference_gpu_duration(text, reference_wav, cfg, steps, speed) -> int:
+    del reference_wav, cfg, steps, speed
+    return _gpu_lease_seconds(text)
 
 try:
     import spaces
 
-    interactive_gpu = spaces.GPU(duration=INTERACTIVE_GPU_LEASE_SECONDS)
-    longform_gpu = spaces.GPU(duration=LONGFORM_GPU_LEASE_SECONDS)
+    speaker_gpu = spaces.GPU(duration=_speaker_gpu_duration)
+    reference_gpu = spaces.GPU(duration=_reference_gpu_duration)
     USING_ZERO_GPU = True
 except ImportError:
-    def interactive_gpu(function):
+    def speaker_gpu(function):
         return function
 
-    longform_gpu = interactive_gpu
+    reference_gpu = speaker_gpu
     USING_ZERO_GPU = False
 
 import gradio as gr
@@ -189,7 +210,7 @@ MIN_ENDPOINT_CUE_UNITS = 6
 SHORT_TEXT_CFG_MIN = 3.0
 SHORT_TEXT_CFG_UNITS = 6
 NETWORK_TEXT_CFG_MIN = 3.0
-CHUNK_CHARS = 80
+CHUNK_CHARS = AUTO_LONG_TEXT_CHARS
 ONSET_CLAUSE_SEARCH_CHARS = 0
 MIN_CHUNK_CHARS = 12
 QUALITY_MEDIUM_TEXT_MAX_UNITS = 48
@@ -3789,7 +3810,7 @@ def _maybe_synthesize_glyph_speaker_request(
         ) from error
 
 
-@interactive_gpu
+@speaker_gpu
 @timed_latency_request
 def tts_speaker(
     text: str,
@@ -3815,7 +3836,7 @@ def tts_speaker(
     )
 
 
-@interactive_gpu
+@reference_gpu
 @timed_latency_request
 def tts_reference(
     text: str,
@@ -3838,46 +3859,14 @@ def tts_reference(
         )
     )
 
-
-@longform_gpu
-@timed_latency_request
-def tts_longform(
-    text: str,
-    speaker: str = DEFAULT_SPEAKER,
-    cfg: float = DEFAULT_CFG,
-    steps: int = DEFAULT_STEPS,
-    speed: float = 1.0,
-):
-    return pcm16_audio_output(
-        *_synthesize(
-            text,
-            SPEAKERS.get(speaker, DEFAULT_CENTROID),
-            cfg=cfg,
-            steps=steps,
-            speed=speed,
-            request_seed=SPEAKER_GENERATION_SEED,
-            speaker_projector_scale=(
-                SPEAKER_B_PROJECTOR_SCALE
-                if speaker == SPEAKER_B_LABEL
-                else 1.0
-            ),
-        )
-    )
-
-
 EXAMPLE_TEXTS = [
     "今天天氣真好，我們一起去散步吧。",
     "我要吃蚵仔煎，然後去丟垃圾。",
     "這學期的成績包括研究報告和期末考。",
     "這是 AI TTS code switching 測試，混合中英文也沒問題。",
     "注音符號測試：ㄅ、ㄆ、ㄇ、ㄈ。",
-]
-
-LONGFORM_EXAMPLES = [
     "今天的會議會先整理目前進度，再確認下一階段的工作。遇到需要討論的項目時，"
     "請先記下問題，等報告結束後再一起處理。最後，我們會確認負責人和預計完成時間。",
-    "歡迎收聽今天的內容。第一段會介紹背景，第二段整理實際案例，最後一段則說明後續安排。"
-    "如果中途聽到英文術語，不必擔心，我們會用中文補充它的意思。",
 ]
 
 
@@ -3947,25 +3936,6 @@ with gr.Blocks(title="BlueMagpie-TTS Demo", theme=gr.themes.Soft()) as demo:
             reference_output,
         )
 
-    with gr.Tab("長文"):
-        with gr.Row():
-            with gr.Column():
-                longform_speaker = gr.Dropdown(list(SPEAKERS), value=DEFAULT_SPEAKER, label="語者")
-                longform_text = gr.Textbox(
-                    label="長文（自動分段）",
-                    lines=8,
-                    max_lines=12,
-                )
-                longform_button = gr.Button("合成完整長文", variant="primary")
-            with gr.Column():
-                longform_output = gr.Audio(label="合成結果", type="numpy")
-        gr.Examples(LONGFORM_EXAMPLES, inputs=longform_text, label="長文範例")
-        longform_button.click(
-            tts_longform,
-            [longform_text, longform_speaker, cfg_input, steps_input, speed_input],
-            longform_output,
-        )
-
     gr.Markdown(
         "合成語音僅供研究與評估展示；正式使用前請人工檢視。 "
         "[模型](https://huggingface.co/OpenFormosa/BlueMagpie-TTS) · "

tests/test_latency_first_mode.py

@@ -88,7 +88,7 @@ def test_public_hot_path_contains_no_quality_gate_or_retry_cascade():
         "_generate_fast_chunk",
         "_assemble_fast_audio",
         "tts_speaker",
-        "tts_longform",
+        "tts_reference",
     ):
         called = {
             node.func.id
@@ -129,7 +129,6 @@ def test_builtin_speaker_a_is_the_default():
     source = APP_PATH.read_text(encoding="utf-8")
     load_speakers = ast.get_source_segment(source, _function_node("_load_speakers"))
     speaker_source = ast.get_source_segment(source, _function_node("tts_speaker"))
-    longform_source = ast.get_source_segment(source, _function_node("tts_longform"))
 
     assert load_speakers is not None
     assert "default_index = 0" in load_speakers
@@ -145,10 +144,23 @@ def test_builtin_speaker_a_is_the_default():
     assert "request_seed=SPEAKER_GENERATION_SEED" in speaker_source
     assert "speaker_projector_scale=(" in speaker_source
     assert "if speaker == SPEAKER_B_LABEL" in speaker_source
-    assert longform_source is not None
-    assert "request_seed=SPEAKER_GENERATION_SEED" in longform_source
-    assert "speaker_projector_scale=(" in longform_source
-    assert "if speaker == SPEAKER_B_LABEL" in longform_source
+
+
+def test_speaker_and_reference_modes_automatically_reserve_long_text_lease():
+    source = APP_PATH.read_text(encoding="utf-8")
+    lease = _isolated_function(
+        "_gpu_lease_seconds",
+        {
+            "AUTO_LONG_TEXT_CHARS": 80,
+            "INTERACTIVE_GPU_LEASE_SECONDS": 60,
+            "LONGFORM_GPU_LEASE_SECONDS": 120,
+        },
+    )
+
+    assert lease("甲" * 80) == 60
+    assert lease("甲" * 81) == 120
+    assert "def tts_longform(" not in source
+    assert 'with gr.Tab("長文")' not in source
 
 
 def test_builtin_speaker_b_uses_the_pinned_checkpoint_embedding(tmp_path):

tests/test_latency_timing.py

@@ -204,9 +204,8 @@ def test_space_wires_every_requested_latency_stage_without_gate_changes():
     assert _decorator_stage(echo_functions["evaluate_echo_smearing"]) == "echo"
 
     expected_gpu_decorators = {
-        "tts_speaker": "interactive_gpu",
-        "tts_reference": "interactive_gpu",
-        "tts_longform": "longform_gpu",
+        "tts_speaker": "speaker_gpu",
+        "tts_reference": "reference_gpu",
     }
     for name, gpu_decorator in expected_gpu_decorators.items():
         decorators = {

tests/test_release_pins.py

@@ -998,7 +998,7 @@ def test_internal_synthesize_accepts_only_a_keyword_seed_while_ui_stays_unchange
     assert synthesize.args.kw_defaults[-1].value is None
     assert "request_seed = resolve_request_seed(request_seed, secrets.randbelow)" in source
     assert "run_coverage_adaptive_cascade(\n                chunks,\n                request_seed," in source
-    for wrapper_name in ("tts_speaker", "tts_reference", "tts_longform"):
+    for wrapper_name in ("tts_speaker", "tts_reference"):
         wrapper = functions[wrapper_name]
         arguments = wrapper.args.args + wrapper.args.kwonlyargs
         assert "request_seed" not in [argument.arg for argument in arguments]
@@ -1155,7 +1155,7 @@ def test_public_tts_wrappers_serialize_the_same_pcm16_grid_used_for_verification
     assert "return pcm16_verification_waveform(waveform)" in assemble_source
     assert "pcm16_audio_output" not in synthesize_source
     assert "return SR, waveform" in synthesize_source
-    for wrapper_name in ("tts_speaker", "tts_reference", "tts_longform"):
+    for wrapper_name in ("tts_speaker", "tts_reference"):
         wrapper_source = ast.get_source_segment(source, functions[wrapper_name])
         assert wrapper_source is not None
         assert "return pcm16_audio_output(" in wrapper_source
@@ -1421,22 +1421,23 @@ def test_space_disables_experimental_gradio_ssr():
     )
 
 
-def test_space_uses_shorter_interactive_zero_gpu_lease():
+def test_space_automatically_selects_zero_gpu_lease_from_text_length():
     source = (ROOT / "app.py").read_text(encoding="utf-8")
 
     assert "INTERACTIVE_GPU_LEASE_SECONDS = 60" in source
     assert "LONGFORM_GPU_LEASE_SECONDS = 120" in source
+    assert "AUTO_LONG_TEXT_CHARS = 80" in source
     assert (
-        "interactive_gpu = spaces.GPU("
-        "duration=INTERACTIVE_GPU_LEASE_SECONDS)"
+        "speaker_gpu = spaces.GPU(duration=_speaker_gpu_duration)"
         in source
     )
     assert (
-        "longform_gpu = spaces.GPU(duration=LONGFORM_GPU_LEASE_SECONDS)"
+        "reference_gpu = spaces.GPU(duration=_reference_gpu_duration)"
         in source
     )
-    assert source.count("@interactive_gpu") == 2
-    assert source.count("@longform_gpu") == 1
+    assert source.count("@speaker_gpu") == 1
+    assert source.count("@reference_gpu") == 1
+    assert "def tts_longform(" not in source
 
 
 def test_network_fragment_relaxation_is_range_bound_and_local_only():