Create app.py

app.py

@@ -0,0 +1,290 @@
+"""
+PlotweaverNigerianVoice
+A Gradio Space for the PlotweaverAI Nigerian-English fine-tuned F5-TTS model.
+
+Two modes:
+  1. Default voice: type text -> generated speech using the Nigerian-English
+     voice baked into this Space (sample.wav / sample.txt from the model repo).
+  2. Custom voice clone: upload your own short reference clip (+ transcript,
+     or leave blank to auto-transcribe) -> generated speech in that voice.
+
+Note: F5-TTS's own infer pipeline already auto-transcribes (Whisper, via
+transformers) and auto-trims reference audio when ref_text is left blank,
+so we lean on that built-in behavior rather than duplicating it here.
+
+Model: PlotweaverAI/nigerian-english-ft-tts (private HF model repo)
+Base architecture: F5-TTS (SWivid/F5-TTS)
+"""
+
+import os
+import threading
+
+import gradio as gr
+import torch
+from huggingface_hub import hf_hub_download
+
+# --- Optional: only present when running on HF Spaces with a GPU tier ----
+try:
+    import spaces
+
+    ON_SPACES = True
+except ImportError:
+    ON_SPACES = False
+
+
+def gpu_decorator(func):
+    """No-op on CPU Spaces; enables ZeroGPU/queued GPU access if upgraded later."""
+    if ON_SPACES:
+        return spaces.GPU(func)
+    return func
+
+
+# F5-TTS imports (package: f5-tts, installed from PyPI / git in requirements.txt)
+from f5_tts.api import F5TTS
+
+
+MODEL_REPO = "PlotweaverAI/nigerian-english-ft-tts"
+HF_TOKEN = os.environ.get("HF_TOKEN")  # set as a Space secret (repo is private)
+
+# Architecture the checkpoint was fine-tuned from. F5TTS_v1_Base is the current
+# default for new finetunes; override with the F5TTS_ARCH secret/variable if
+# your training run used a different base (e.g. "F5TTS_Base").
+MODEL_ARCH = os.environ.get("F5TTS_ARCH", "F5TTS_v1_Base")
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+MAX_GEN_CHARS = 600          # keep generations bounded on CPU
+
+
+_model_lock = threading.Lock()
+_tts_model = None
+_default_ref_audio = None
+_default_ref_text = None
+
+
+def _download_model_files():
+    """Pull the fine-tuned checkpoint, vocab, and bundled sample voice
+    from the private model repo. Requires HF_TOKEN secret with read access."""
+    ckpt_path = hf_hub_download(
+        repo_id=MODEL_REPO, filename="model_last.pt", token=HF_TOKEN
+    )
+    vocab_path = hf_hub_download(
+        repo_id=MODEL_REPO, filename="vocab.txt", token=HF_TOKEN
+    )
+    sample_wav_path = hf_hub_download(
+        repo_id=MODEL_REPO, filename="sample.wav", token=HF_TOKEN
+    )
+    sample_txt_path = hf_hub_download(
+        repo_id=MODEL_REPO, filename="sample.txt", token=HF_TOKEN
+    )
+    with open(sample_txt_path, "r", encoding="utf-8") as f:
+        sample_text = f.read().strip()
+    return ckpt_path, vocab_path, sample_wav_path, sample_text
+
+
+def get_model():
+    """Lazily load the F5-TTS model + default reference voice exactly once."""
+    global _tts_model, _default_ref_audio, _default_ref_text
+    if _tts_model is not None:
+        return _tts_model, _default_ref_audio, _default_ref_text
+
+    with _model_lock:
+        if _tts_model is not None:
+            return _tts_model, _default_ref_audio, _default_ref_text
+
+        ckpt_path, vocab_path, sample_wav_path, sample_text = _download_model_files()
+
+        try:
+            model = F5TTS(
+                model=MODEL_ARCH,
+                ckpt_file=ckpt_path,
+                vocab_file=vocab_path,
+                device=DEVICE,
+            )
+        except RuntimeError as e:
+            raise RuntimeError(
+                f"Failed to load checkpoint with architecture '{MODEL_ARCH}'. "
+                "If this fine-tune was trained from a different F5-TTS base "
+                "(e.g. 'F5TTS_Base' instead of 'F5TTS_v1_Base'), set the "
+                "F5TTS_ARCH variable in your Space settings to match. "
+                f"Original error: {e}"
+            ) from e
+
+        _tts_model = model
+        _default_ref_audio = sample_wav_path
+        _default_ref_text = sample_text
+
+    return _tts_model, _default_ref_audio, _default_ref_text
+
+
+@gpu_decorator
+def generate_default_voice(text: str, speed: float, nfe_steps: int):
+    if not text or not text.strip():
+        raise gr.Error("Please enter some text to generate speech for.")
+    if len(text) > MAX_GEN_CHARS:
+        raise gr.Error(
+            f"Text is too long ({len(text)} characters). "
+            f"Please keep it under {MAX_GEN_CHARS} characters on this CPU Space."
+        )
+
+    model, ref_audio, ref_text = get_model()
+
+    wav, sr, _ = model.infer(
+        ref_file=ref_audio,
+        ref_text=ref_text,
+        gen_text=text.strip(),
+        speed=speed,
+        nfe_step=int(nfe_steps),
+        remove_silence=True,
+    )
+
+    return (sr, wav)
+
+
+@gpu_decorator
+def generate_cloned_voice(
+    ref_audio_path: str,
+    ref_text: str,
+    gen_text: str,
+    speed: float,
+    nfe_steps: int,
+):
+    if ref_audio_path is None:
+        raise gr.Error("Please upload a reference voice clip first.")
+    if not gen_text or not gen_text.strip():
+        raise gr.Error("Please enter the text you want spoken in the cloned voice.")
+    if len(gen_text) > MAX_GEN_CHARS:
+        raise gr.Error(
+            f"Text is too long ({len(gen_text)} characters). "
+            f"Please keep it under {MAX_GEN_CHARS} characters on this CPU Space."
+        )
+
+    model, _, _ = get_model()
+
+    transcript = ref_text.strip() if ref_text else ""
+    if not transcript:
+        # F5-TTS's transcribe() uses a Whisper pipeline under the hood.
+        transcript = model.transcribe(ref_audio_path)
+        if not transcript:
+            raise gr.Error(
+                "Could not auto-transcribe the uploaded clip. "
+                "Please type the transcript manually and try again."
+            )
+
+    wav, sr, _ = model.infer(
+        ref_file=ref_audio_path,
+        ref_text=transcript,
+        gen_text=gen_text.strip(),
+        speed=speed,
+        nfe_step=int(nfe_steps),
+        remove_silence=True,
+    )
+
+    return (sr, wav), transcript
+
+CSS = """
+#title { text-align: center; margin-bottom: 0.5em; }
+#subtitle { text-align: center; color: var(--body-text-color-subdued); margin-bottom: 1.5em; }
+.cpu-note { font-size: 0.85em; color: var(--body-text-color-subdued); }
+"""
+
+with gr.Blocks(css=CSS, title="Plotweaver Nigerian Voice") as demo:
+    gr.Markdown("#Plotweaver Nigerian Voice", elem_id="title")
+    gr.Markdown(
+        "Nigerian-English text-to-speech, fine-tuned from F5-TTS by Plotweaver AI. "
+        "Generate speech in our Nigerian voice, or clone a voice from your own clip.",
+        elem_id="subtitle",
+    )
+    gr.Markdown(
+        "Running on free CPU hardware — generation can take **30–90+ seconds** "
+        "per request, longer for longer text. Research / non-commercial use only "
+        "(base F5-TTS is CC-BY-NC).",
+        elem_classes="cpu-note",
+    )
+
+    with gr.Tabs():
+        
+        with gr.Tab("Nigerian Voice (Text → Speech)"):
+            with gr.Row():
+                with gr.Column(scale=1):
+                    default_text = gr.Textbox(
+                        label="Text to speak",
+                        placeholder="Type what you want the Nigerian voice to say...",
+                        lines=6,
+                        max_lines=12,
+                    )
+                    with gr.Accordion("Advanced settings", open=False):
+                        default_speed = gr.Slider(
+                            0.5, 2.0, value=1.0, step=0.05, label="Speed"
+                        )
+                        default_nfe = gr.Slider(
+                            8, 64, value=24, step=2,
+                            label="Quality steps (NFE)",
+                            info="Higher = better quality but slower. 16-24 recommended on CPU.",
+                        )
+                    default_btn = gr.Button("Generate Speech", variant="primary")
+                with gr.Column(scale=1):
+                    default_audio_out = gr.Audio(label="Generated audio", type="numpy")
+
+            default_btn.click(
+                fn=generate_default_voice,
+                inputs=[default_text, default_speed, default_nfe],
+                outputs=[default_audio_out],
+            )
+
+        with gr.Tab("Clone a Voice (Upload Clip → Speech)"):
+            gr.Markdown(
+                "Upload a short, clean voice clip (5-15 seconds works best). "
+                "Add the transcript of that clip if you have it — or leave it "
+                "blank and we'll auto-transcribe it with Whisper."
+            )
+            with gr.Row():
+                with gr.Column(scale=1):
+                    ref_audio_in = gr.Audio(
+                        label="Reference voice clip",
+                        type="filepath",
+                        sources=["upload", "microphone"],
+                    )
+                    ref_text_in = gr.Textbox(
+                        label="Transcript of the clip (optional — auto-transcribed if left blank)",
+                        placeholder="Leave blank to auto-transcribe with Whisper...",
+                        lines=3,
+                    )
+                    clone_gen_text = gr.Textbox(
+                        label="Text to speak in this cloned voice",
+                        placeholder="Type what you want spoken in the uploaded voice...",
+                        lines=5,
+                        max_lines=12,
+                    )
+                    with gr.Accordion("Advanced settings", open=False):
+                        clone_speed = gr.Slider(
+                            0.5, 2.0, value=1.0, step=0.05, label="Speed"
+                        )
+                        clone_nfe = gr.Slider(
+                            8, 64, value=24, step=2,
+                            label="Quality steps (NFE)",
+                            info="Higher = better quality but slower. 16-24 recommended on CPU.",
+                        )
+                    clone_btn = gr.Button("Generate Cloned Speech", variant="primary")
+                with gr.Column(scale=1):
+                    clone_audio_out = gr.Audio(label="Generated audio", type="numpy")
+                    used_transcript_out = gr.Textbox(
+                        label="Transcript used for the reference clip",
+                        interactive=False,
+                    )
+
+            clone_btn.click(
+                fn=generate_cloned_voice,
+                inputs=[ref_audio_in, ref_text_in, clone_gen_text, clone_speed, clone_nfe],
+                outputs=[clone_audio_out, used_transcript_out],
+            )
+
+    gr.Markdown(
+        "---\nBuilt on [F5-TTS](https://github.com/SWivid/F5-TTS) "
+        "(CC-BY-NC license) · Fine-tuned voice by Plotweaver AI · "
+        "Non-commercial / research use.",
+        elem_classes="cpu-note",
+    )
+
+if __name__ == "__main__":
+    demo.launch()