v1: HF Space frontend + Modal GPU backend + codemix recipe

.gitattributes

@@ -33,3 +33,7 @@ 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
+data/cartesia_hi_female_6s.wav filter=lfs diff=lfs merge=lfs -text
+data/sarvam_hi_female_10s.wav filter=lfs diff=lfs merge=lfs -text
+data/sarvam_ta_male_11s.wav filter=lfs diff=lfs merge=lfs -text
+data/sarvam_te_female_9s.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,12 +1,69 @@
 ---
-title: Praxy Voice Demo
-emoji: 📊
-colorFrom: red
+title: Praxy Voice
+emoji: 🎤
+colorFrom: indigo
 colorTo: purple
 sdk: gradio
-sdk_version: 6.13.0
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
+license: apache-2.0
+short_description: Hi · Te · Ta · En TTS · code-mix · voice cloning
+models:
+  - Praxel/praxy-voice-r6
+  - ResembleAI/chatterbox
+  - ai4bharat/IndicF5
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+## Required Space secrets
+- `MODAL_TOKEN_ID` — get from Modal dashboard
+- `MODAL_TOKEN_SECRET`
+- `ANTHROPIC_API_KEY` — for the codemix transliteration preprocessor
+
+## Hardware
+Free CPU is sufficient — this Space dispatches all GPU work to Modal.
+
+## Custom domain (voice.praxel.in)
+HF Space settings → Custom Domain → add `voice.praxel.in`, then add CNAME at praxel.in DNS pointing to the Space URL.
+
+# Praxy Voice — Gradio Demo
+
+Live demo of **Praxy Voice**, an open-source commercial-class Indic TTS system. Type Telugu, Tamil, or Hindi text and hear it spoken in a natural voice — with your own uploaded reference clip, or a bundled default.
+
+## How it works
+
+Three branches behind a single interface:
+
+| Language | Path |
+|---|---|
+| Telugu | R6 LoRA on Chatterbox + BUPS romanisation + Hi-proxy `language_id` |
+| Tamil  | R6 LoRA on Chatterbox + BUPS romanisation + Hi-proxy `language_id` |
+| Hindi  | **Vanilla** Chatterbox (LoRA not needed — Chatterbox natively covers Hi) |
+
+All three share the **voice-prompt recovery** recipe: an 8–11 s same-language reference audio clip + Config B sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1).
+
+See the companion paper for the full methodology: *"Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS"* (arXiv 2026).
+
+## Local run
+
+```bash
+cd gradio_demo
+pip install -r requirements.txt
+python app.py
+```
+
+## Files
+
+- `app.py` — Gradio interface + language router + inference
+- `requirements.txt` — Python deps
+- `data/` — bundled default reference clips (place your own here before deploy)
+
+## Defaults and limits
+
+- Default reference clips: Sarvam Bulbul Te/Ta + Cartesia Sonic-3 Hi (8–11 s each).
+- Max input text length: not enforced in UI but practically ~400 chars.
+- Hardware: runs on ZeroGPU (T4-class; ~2–5 s per generation).
+
+## License
+
+Apache-2.0. See `LICENSE` in the model repo `Praxel/praxy-voice-r6`.

app.py

@@ -0,0 +1,298 @@
+"""Praxy Voice — Gradio app on HF Space, GPU inference on Modal.
+
+Live at:
+    https://huggingface.co/spaces/Praxel/praxy-voice-demo
+    https://voice.praxel.in (CNAME → HF Space, see README)
+
+Architecture: this Space runs the Gradio UI on HF's free CPU tier and
+dispatches synth requests cross-cluster to the deployed `praxy-voice`
+Modal app's classes (PraxyChatterboxLoRA + IndicF5TTS). Modal has the
+GPU + the proven IndicF5 dependency chain; HF Space is the right tool
+for serving Gradio static assets (which the Modal asgi_app proxy
+doesn't handle cleanly — see memory/project_modal_gradio_dead_end_2026-04-27.md).
+
+Required HF Space secrets:
+    MODAL_TOKEN_ID
+    MODAL_TOKEN_SECRET
+    ANTHROPIC_API_KEY     # for the codemix transliteration preprocessor
+
+Required HF Space hardware: free CPU is fine; this app does no GPU work.
+
+Sync from repo (run from `/Users/pushpak/Documents/GitHub/praxy_tts`):
+    huggingface-cli upload Praxel/praxy-voice-demo gradio_demo/ . --repo-type=space
+
+Or push via git remote:
+    git remote add space https://huggingface.co/spaces/Praxel/praxy-voice-demo
+    git subtree push --prefix gradio_demo space main
+"""
+
+from __future__ import annotations
+
+import io
+import os
+import re
+import tempfile
+import time
+from pathlib import Path
+
+import gradio as gr
+import modal
+
+HERE = Path(__file__).resolve().parent
+
+# Voice library — pre-bundled reference voices.
+VOICE_LIBRARY: dict[str, dict] = {
+    "Hindi · Female (default)": {
+        "lang": "hi",
+        "ref_audio": str(HERE / "data" / "sarvam_hi_female_10s.wav"),
+        "ref_text": "मेरे दादा जी हर शाम बरामदे में बैठकर अपने बचपन की कहानियां सुनाते हैं और हम सब मिलकर बडे चाव से उनकी बातें सुनते हैं",
+    },
+    "Hindi · Female (alt)": {
+        "lang": "hi",
+        "ref_audio": str(HERE / "data" / "cartesia_hi_female_6s.wav"),
+        "ref_text": "मेरे दादाजी हर शाम बरामदे में बैठकर अपने बचपन की कहानियां सुनाते हैं और हम सब मिलकर बड़े चाव से उनकी बातें सुनते हैं",
+    },
+    "Telugu · Female": {
+        "lang": "te",
+        "ref_audio": str(HERE / "data" / "sarvam_te_female_9s.wav"),
+        "ref_text": "మా తాతయ్య ప్రతి సాయంత్రం వరండాలో కూర్చుని తన చిన్నతనంలో జరిగిన కథలు చెబుతూ ఉంటారు మరియు మేము అందరూ కలిసి ఆసక్తిగా వినేవాళ్లం",
+    },
+    "Tamil · Male": {
+        "lang": "ta",
+        "ref_audio": str(HERE / "data" / "sarvam_ta_male_11s.wav"),
+        "ref_text": "எங்கள் தாத்தா தினமும் மாலையில் திண்ணையில் அமர்ந்து கொண்டு தன் சிறுவயதில் நடந்த கதைகளைச் சொல்லிக் கொண்டிருப்பார் நாங்கள் அனைவரும் சேர்ந்து ஆர்வமாகக் கேட்போம்",
+    },
+}
+
+LANG_CODE_TO_NAME = {"hi": "Hindi", "te": "Telugu", "ta": "Tamil", "en": "English"}
+
+EXAMPLES = [
+    ["Hindi · Female (default)", "नमस्ते! मेरा नाम प्राक्सी है, और मैं भारत के लिए बनाई गई एक ओपन-सोर्स आवाज़ हूँ।"],
+    ["Hindi · Female (default)", "मैंने WhatsApp पे message किया but notification नहीं आया।"],
+    ["Telugu · Female", "నమస్తే! నేను ప్రాక్సీ, ఇండియా కోసం రూపొందించిన ఓపెన్-సోర్స్ వాయిస్."],
+    ["Telugu · Female", "మా CEO ఇవాళ all-hands meeting లో కొత్త quarterly targets announce చేశారు."],
+    ["Tamil · Male", "வணக்கம்! நான் ப்ராக்ஸி, இந்தியாவுக்காக உருவாக்கப்பட்ட ஓபன்-சோர்ஸ் குரல்."],
+]
+
+# Cross-cluster Modal lookups — happen lazily on first synth call so the
+# Space boots fast even if Modal is slow.
+_indicf5 = None
+_chatterbox = None
+
+
+def _get_modal_handles():
+    global _indicf5, _chatterbox
+    if _indicf5 is None:
+        _indicf5 = modal.Cls.from_name("praxy-voice", "IndicF5TTS")()
+    if _chatterbox is None:
+        _chatterbox = modal.Cls.from_name("praxy-voice", "PraxyChatterboxLoRA")
+    return _indicf5, _chatterbox
+
+
+def _detect_lang(text: str) -> str:
+    if re.search(r"[ఀ-౿]", text):
+        return "te"
+    if re.search(r"[஀-௿]", text):
+        return "ta"
+    if re.search(r"[ऀ-ॿ]", text):
+        return "hi"
+    return "en"
+
+
+def _is_codemix(text: str, lang: str) -> bool:
+    return bool(re.search(r"[A-Za-z]{2,}", text)) and lang != "en"
+
+
+# Codemix transliteration via Anthropic — inline because we don't want to
+# depend on the `serving.codemix_to_native_script` module on the HF Space
+# (which would require pulling in the whole repo). Same prompt as the
+# canonical implementation.
+LANG_NATIVE_NAMES = {"hi": "Hindi (Devanagari script)", "te": "Telugu", "ta": "Tamil"}
+
+_TRANSLIT_SYSTEM = """You convert Indian code-mix sentences into pure native-script form.
+
+Input: a sentence mixing {lang_name} (in {lang_name} script) with Latin-script English words/phrases.
+
+Output rules:
+1. Keep every {lang_name}-script word/character unchanged. Do NOT translate them.
+2. For every Latin-script English word/phrase, write its phonetic spelling in {lang_name} script — exactly the way an educated native {lang_name} speaker would write it casually (the way Bollywood subtitles, Indian news tickers, and Sarvam-Bulbul's training data spells English brand and tech terms). Examples for Hindi: WhatsApp → व्हाट्सऐप, message → मैसेज, notification → नोटिफिकेशन, CEO → सीईओ, syllabus → सिलेबस, complete → कम्प्लीट, traffic jam → ट्रैफिक जैम, weekend → वीकेंड. For Telugu: WhatsApp → వాట్సాప్, message → మెసేజ్, notification → నోటిఫికేషన్, CEO → సీఈఓ.
+3. Preserve all word order, spacing, and punctuation exactly as in the input.
+4. Do NOT add explanations, brackets, alternatives, or commentary. Output ONLY the converted sentence.
+
+Examples:
+Input (Hindi codemix): मैंने WhatsApp पे message किया but notification नहीं आया।
+Output: मैंने व्हाट्सऐप पे मैसेज किया बट नोटिफिकेशन नहीं आया।
+
+Input (Telugu codemix): మా CEO ఇవాళ all-hands meeting లో కొత్త quarterly targets announce చేశారు.
+Output: మా సీఈఓ ఇవాళ ఆల్-హ్యాండ్స్ మీటింగ్ లో కొత్త క్వార్టర్లీ టార్గెట్స్ అనౌన్స్ చేశారు."""
+
+
+def _transliterate(text: str, lang: str) -> str:
+    if not re.search(r"[A-Za-z]", text):
+        return text
+    import anthropic
+    client = anthropic.Anthropic()  # picks up ANTHROPIC_API_KEY from env
+    sys_prompt = _TRANSLIT_SYSTEM.format(lang_name=LANG_NATIVE_NAMES.get(lang, lang))
+    resp = client.messages.create(
+        model="claude-haiku-4-5",
+        max_tokens=512,
+        temperature=0.0,
+        system=sys_prompt,
+        messages=[{"role": "user", "content": text}],
+    )
+    out = resp.content[0].text.strip()
+    out = re.sub(r"^(Output|Translation|Result):\s*", "", out, flags=re.IGNORECASE).strip()
+    return out
+
+
+def synth(text: str, voice_choice: str, custom_audio_path, custom_text: str):
+    if not text or not text.strip():
+        return None, "❌ Please type some text to synthesise."
+
+    custom_mode = voice_choice.startswith("📤")
+    if custom_mode and custom_audio_path is None:
+        return None, "❌ Pick a pre-made voice or upload a reference clip."
+
+    if custom_mode:
+        ref_audio_path = custom_audio_path
+        ref_text = (custom_text or "").strip()
+        if not ref_text:
+            return None, "❌ Custom voice needs a reference transcript matching the audio."
+        lang = _detect_lang(text)
+        with open(ref_audio_path, "rb") as f:
+            ref_bytes = f.read()
+    else:
+        voice = VOICE_LIBRARY[voice_choice]
+        with open(voice["ref_audio"], "rb") as f:
+            ref_bytes = f.read()
+        ref_text = voice["ref_text"]
+        lang = voice["lang"]
+
+    codemix = _is_codemix(text, lang)
+    indicf5, chatterbox_cls = _get_modal_handles()
+
+    # 1. Codemix → transliterate → IndicF5
+    if codemix:
+        try:
+            text_for_synth = _transliterate(text, lang)
+        except Exception as e:
+            return None, f"❌ Transliteration failed: {type(e).__name__}: {e}"
+        t0 = time.time()
+        try:
+            wav_bytes, _ = indicf5.synthesize.remote(
+                text=text_for_synth, ref_audio_bytes=ref_bytes, ref_text=ref_text,
+            )
+        except Exception as e:
+            return None, f"❌ IndicF5 synth failed: {type(e).__name__}: {e}"
+        out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        out.write(wav_bytes); out.close()
+        return out.name, (
+            f"✓ {LANG_CODE_TO_NAME.get(lang, lang)} code-mix · "
+            f"transliterate→IndicF5 · {time.time()-t0:.1f}s"
+        )
+
+    # 2. Pure Hi or Ta → IndicF5 zero-shot
+    if lang in ("hi", "ta"):
+        t0 = time.time()
+        try:
+            wav_bytes, _ = indicf5.synthesize.remote(
+                text=text, ref_audio_bytes=ref_bytes, ref_text=ref_text,
+            )
+        except Exception as e:
+            return None, f"❌ IndicF5 synth failed: {type(e).__name__}: {e}"
+        out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        out.write(wav_bytes); out.close()
+        return out.name, (
+            f"✓ {LANG_CODE_TO_NAME.get(lang, lang)} pure · "
+            f"IndicF5 zero-shot · {time.time()-t0:.1f}s"
+        )
+
+    # 3. Te pure or English → Chatterbox + R6 LoRA
+    env = {
+        "PRAXY_CKPT_PATH": "/cache/chatterbox_indic/round_6/step_8000.ckpt",
+        "PRAXY_USE_BUPS": "1" if lang == "te" else "0",
+        "PRAXY_NO_LORA": "1" if lang == "en" else "0",
+    }
+    synth_cls = chatterbox_cls.with_options(
+        secrets=[modal.Secret.from_dict(env), modal.Secret.from_name("praxy-hf")]
+    )()
+    t0 = time.time()
+    try:
+        wav_bytes, _ = synth_cls.synthesize.remote(
+            text=text, language_code=lang, ref_audio_bytes=ref_bytes,
+            exaggeration=0.7, cfg_weight=0.5, temperature=0.6,
+            repetition_penalty=2.0, min_p=0.1, top_p=1.0,
+        )
+    except Exception as e:
+        return None, f"❌ Chatterbox synth failed: {type(e).__name__}: {e}"
+    out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+    out.write(wav_bytes); out.close()
+    branch = "vanilla Chatterbox" if lang == "en" else "Chatterbox + R6 LoRA"
+    return out.name, (
+        f"✓ {LANG_CODE_TO_NAME.get(lang, lang)} · {branch} · {time.time()-t0:.1f}s"
+    )
+
+
+CSS = """
+.gradio-container { max-width: 980px !important; }
+h1 { font-weight: 700; margin-bottom: 0.4em; }
+footer { display: none !important; }
+"""
+
+with gr.Blocks(title="Praxy Voice", theme=gr.themes.Soft(), css=CSS) as demo:
+    gr.Markdown(
+        "# 🎤 Praxy Voice\n"
+        "Open-source Hindi · Telugu · Tamil · English TTS, including code-mix. "
+        "Voice cloning from any 8–15 s reference clip. "
+        "Built on [Chatterbox](https://github.com/resemble-ai/chatterbox) + "
+        "[IndicF5](https://huggingface.co/ai4bharat/IndicF5) + a Haiku-driven "
+        "native-script transliteration preprocessor.\n\n"
+        "*v1 · Code: [github.com/praxelhq/praxy-tts](https://github.com/praxelhq/praxy-tts) · "
+        "Paper: arXiv (link soon).*"
+    )
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            text_in = gr.Textbox(
+                label="Text to synthesise",
+                placeholder="मैंने WhatsApp पे message किया but notification नहीं आया।",
+                lines=4,
+            )
+            voice_in = gr.Dropdown(
+                list(VOICE_LIBRARY.keys()) + ["📤 Use custom upload below"],
+                value=list(VOICE_LIBRARY.keys())[0],
+                label="Voice",
+            )
+            with gr.Accordion("📤 Or clone your own voice (8–15 s clip)", open=False):
+                custom_audio = gr.Audio(label="Reference audio", type="filepath")
+                custom_text = gr.Textbox(
+                    label="Reference transcript (must match the audio exactly)",
+                    lines=2,
+                )
+            btn = gr.Button("Generate", variant="primary", size="lg")
+        with gr.Column(scale=1):
+            audio_out = gr.Audio(label="Output", autoplay=True)
+            status = gr.Markdown("")
+
+    gr.Examples(EXAMPLES, inputs=[voice_in, text_in], label="Try one of these")
+
+    btn.click(
+        synth,
+        inputs=[text_in, voice_in, custom_audio, custom_text],
+        outputs=[audio_out, status],
+    )
+
+    gr.Markdown(
+        "---\n"
+        "**Code-mix** (Latin English words inside Indic text) is auto-detected and "
+        "transliterated to native script before synth — that's why *WhatsApp* gets "
+        "pronounced *vaa-ts-ay-p* rather than American *whats-app*. It matches how "
+        "Indians actually code-switch.\n\n"
+        "**First request** triggers a Modal cold-start (~30 s while the GPU container warms). "
+        "Subsequent requests are sub-5 s.\n\n"
+        "**Privacy**: uploaded reference clips are not logged or stored."
+    )
+
+
+if __name__ == "__main__":
+    demo.queue(max_size=10).launch(server_name="0.0.0.0")

data/cartesia_hi_female_6s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1129c0c56aa8e82a724dbfba93083e21f74c131be5864272922665162a0a6044
+size 305828

data/sarvam_hi_female_10s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55b8cc3d3ad8cf35235b93be3c866dd9f7555a404b4fee12f9a88104814f6021
+size 471084

data/sarvam_ta_male_11s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:871fbd6838ab53e85d6e0657c9fe10c07a7a2b13016cb43fc6d395fd32843c35
+size 540716

data/sarvam_te_female_9s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd5b1f59fb031abd8498069d1790129989b5beb415238c1d41b2bb2390004f30
+size 454700

requirements.txt

@@ -0,0 +1,6 @@
+gradio==4.44.0
+huggingface_hub<1.0
+modal>=0.64
+anthropic>=0.40
+soundfile==0.12.1
+numpy<2