Add KasbahTTS V0 Algerian Dardja demo

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/ALG.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,45 @@
 ---
-title: KasbahTTS Demo
-emoji: 🐨
-colorFrom: purple
-colorTo: indigo
+title: KasbahTTS V0 — Algerian Dardja TTS
+emoji: 🎙️
+colorFrom: red
+colorTo: green
 sdk: gradio
-sdk_version: 6.12.0
-python_version: '3.12'
+sdk_version: 5.50.0
 app_file: app.py
 pinned: false
+license: mit
+short_description: Zero-shot Algerian Dardja (الدارجة) text-to-speech
+models:
+  - MenaVoice/KasbahTTS-V0
+tags:
+  - text-to-speech
+  - arabic
+  - algerian-arabic
+  - dardja
+  - f5-tts
+  - voice-cloning
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# KasbahTTS V0 — Algerian Dardja TTS
+
+Interactive demo for [`MenaVoice/KasbahTTS-V0`](https://huggingface.co/MenaVoice/KasbahTTS-V0),
+the first open-source text-to-speech model built for **Algerian Dardja** (الدارجة الجزائرية).
+
+Give it a few seconds of any voice and some Dardja text, and it speaks that text in that voice.
+
+## How it works
+
+Built on **F5-TTS** (DiT-based flow matching), fine-tuned from **Habibi-TTS** on real Algerian
+conversational speech. Weights are pulled at runtime from the model repo, so this Space does not
+carry its own copy.
+
+## Limitations
+
+- Arabic script only — no French code-switching
+- No digits — spell numbers out (`ثلاثة`, not `3`)
+- No diacritics (تشكيل)
+- May occasionally repeat words; try `nfe_step=64` or shorter segments
+
+---
+
+*من القصبة للعالم — From the Kasbah to the world*

app.py

@@ -0,0 +1,201 @@
+"""
+KasbahTTS V0 — Algerian Dardja text-to-speech.
+
+Zero-shot voice cloning demo for MenaVoice/KasbahTTS-V0, an F5-TTS (DiT flow
+matching) model fine-tuned on Algerian Dardja.
+
+`spaces` must be imported before torch so ZeroGPU can patch CUDA init.
+"""
+
+import os
+import random
+import re
+import tempfile
+
+try:
+    import spaces
+
+    USING_ZEROGPU = True
+except ImportError:
+    USING_ZEROGPU = False
+
+import gradio as gr
+import soundfile as sf
+import torch
+from f5_tts.infer.utils_infer import (
+    load_model,
+    load_vocoder,
+    preprocess_ref_audio_text,
+    remove_silence_for_generated_wav,
+)
+from f5_tts.model import DiT
+from huggingface_hub import hf_hub_download
+
+from habibi_tts.infer.utils_infer import infer_process
+
+
+def gpu_decorator(func):
+    # Only wrap with the ZeroGPU allocator when actually running on ZeroGPU
+    # hardware (env var set by HF). On CPU/paid-GPU spaces this is a no-op.
+    if USING_ZEROGPU and os.environ.get("SPACES_ZERO_GPU"):
+        return spaces.GPU(duration=120)(func)
+    return func
+
+
+MODEL_REPO = "MenaVoice/KasbahTTS-V0"
+CKPT_FILE = "ALGERIA.safetensors"
+V1_CFG = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+ckpt_path = hf_hub_download(MODEL_REPO, CKPT_FILE)
+vocab_path = hf_hub_download(MODEL_REPO, "vocab.txt")
+
+vocoder = load_vocoder(vocoder_name="vocos", is_local=False, device=DEVICE)
+model = load_model(DiT, V1_CFG, ckpt_path, vocab_file=vocab_path, device=DEVICE)
+
+EXAMPLE_REF = os.path.join(os.path.dirname(__file__), "examples", "ALG.wav")
+EXAMPLE_REF_TEXT = "أنيا هكا باغية ناكل هكا أني ن نشوف فيها الحاجة هذيكا."
+
+# The model was trained on unvocalized Arabic script only: Latin letters and
+# digits fall outside its vocabulary and produce unpredictable audio.
+UNSUPPORTED = re.compile(r"[A-Za-z0-9٠-٩]")
+
+
+@gpu_decorator
+def generate(
+    ref_audio_orig,
+    ref_text,
+    gen_text,
+    speed,
+    nfe_step,
+    cfg_strength,
+    cross_fade,
+    remove_sil,
+    seed,
+    progress=gr.Progress(),
+):
+    if not ref_audio_orig:
+        raise gr.Error("Please provide a reference audio clip (5-15 seconds works best).")
+    if not gen_text.strip():
+        raise gr.Error("Please enter some Algerian Dardja text to synthesize.")
+
+    if UNSUPPORTED.search(gen_text):
+        gr.Warning(
+            "Latin letters and digits are outside the model's vocabulary. "
+            "Spell numbers out in Arabic words (ثلاثة, not 3)."
+        )
+
+    if seed is None or int(seed) < 0:
+        seed = random.randint(0, 2**31 - 1)
+    seed = int(seed)
+    torch.manual_seed(seed)
+
+    # Leaving ref_text blank triggers Whisper auto-transcription inside f5-tts.
+    ref_audio, ref_text = preprocess_ref_audio_text(ref_audio_orig, ref_text or "", show_info=gr.Info)
+
+    # KasbahTTS is a *specialized* single-dialect checkpoint, so no dialect token
+    # is prepended — infer_process defaults dialect_id to None.
+    wave, sample_rate, _ = infer_process(
+        ref_audio,
+        ref_text,
+        gen_text,
+        model,
+        vocoder,
+        cross_fade_duration=cross_fade,
+        nfe_step=int(nfe_step),
+        cfg_strength=cfg_strength,
+        speed=speed,
+        show_info=gr.Info,
+        progress=progress,
+    )
+
+    if remove_sil:
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
+            tmp_path = f.name
+        try:
+            sf.write(tmp_path, wave, sample_rate)
+            remove_silence_for_generated_wav(tmp_path)
+            wave, sample_rate = sf.read(tmp_path)
+        finally:
+            os.unlink(tmp_path)
+
+    return (sample_rate, wave), ref_text, seed
+
+
+with gr.Blocks(title="KasbahTTS V0", theme=gr.themes.Soft()) as demo:
+    gr.Markdown(
+        """
+        # 🇩🇿 KasbahTTS V0 — Algerian Dardja TTS
+
+        Zero-shot voice cloning for **Algerian Dardja** (الدارجة الجزائرية).
+        Upload a few seconds of any voice, type Dardja text, and hear it spoken back in that voice.
+
+        Model: [`MenaVoice/KasbahTTS-V0`](https://huggingface.co/MenaVoice/KasbahTTS-V0) ·
+        F5-TTS architecture, fine-tuned from Habibi-TTS · MIT licensed.
+        """
+    )
+
+    with gr.Row():
+        with gr.Column():
+            ref_audio = gr.Audio(label="Reference voice", type="filepath")
+            ref_text = gr.Textbox(
+                label="Reference transcript (optional)",
+                lines=2,
+                placeholder="Leave blank to auto-transcribe with Whisper…",
+            )
+            gen_text = gr.Textbox(
+                label="Text to generate (Arabic script only)",
+                lines=4,
+                placeholder="واش راك خويا، لاباس عليك؟",
+                rtl=True,
+            )
+            run = gr.Button("🎤 Generate", variant="primary", size="lg")
+
+        with gr.Column():
+            audio_out = gr.Audio(label="Generated speech", interactive=False)
+            ref_text_out = gr.Textbox(label="Reference transcript used", interactive=False, lines=2)
+            seed_out = gr.Number(label="Seed used", interactive=False, precision=0)
+
+            with gr.Accordion("Advanced settings", open=False):
+                speed = gr.Slider(0.5, 1.5, value=1.0, step=0.05, label="Speed")
+                nfe_step = gr.Slider(
+                    8, 64, value=32, step=8, label="NFE steps",
+                    info="More steps = higher quality, slower.",
+                )
+                cfg_strength = gr.Slider(0.5, 5.0, value=2.0, step=0.1, label="CFG strength")
+                cross_fade = gr.Slider(0.0, 0.5, value=0.15, step=0.05, label="Cross-fade (s)")
+                remove_sil = gr.Checkbox(label="Trim long silences", value=False)
+                seed = gr.Number(label="Seed (-1 = random)", value=-1, precision=0)
+
+    inputs = [ref_audio, ref_text, gen_text, speed, nfe_step, cfg_strength, cross_fade, remove_sil, seed]
+    outputs = [audio_out, ref_text_out, seed_out]
+
+    run.click(generate, inputs=inputs, outputs=outputs)
+
+    if os.path.exists(EXAMPLE_REF):
+        gr.Examples(
+            examples=[
+                [EXAMPLE_REF, EXAMPLE_REF_TEXT, "واش راك خويا، لاباس عليك؟ صباح الخير."],
+                [EXAMPLE_REF, EXAMPLE_REF_TEXT, "من القصبة للعالم، هذا أول موديل يهدر بالدارجة الجزائرية."],
+            ],
+            inputs=[ref_audio, ref_text, gen_text],
+        )
+
+    gr.Markdown(
+        """
+        ---
+        ### Known limitations
+
+        - **Arabic script only.** No French code-switching — `ça va` will not work.
+        - **No digits.** Write `ثلاثة`, not `3`.
+        - **No diacritics (تشكيل).** Use plain, unvocalized Arabic.
+        - **Occasional repetition.** Try `nfe_step` 64, adjust `cfg_strength`, or split long text.
+
+        Built with ❤️ for Algeria by MenaVoice — *من القصبة للعالم*
+        """
+    )
+
+
+if __name__ == "__main__":
+    demo.queue().launch()

examples/ALG.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:187b5992f7b068371f2769860a8b776e50abcc4168d9d24811ebb97ec0b8664f
+size 294174

requirements.txt

@@ -0,0 +1,3 @@
+habibi-tts==0.1.1
+soundfile
+spaces