Build-Small submission: field-guide UI + Modal backend

README.md

@@ -1,13 +1,47 @@
 ---
 title: Polyglot Me
-emoji: 🏆
-colorFrom: yellow
+emoji: 🎙️
+colorFrom: indigo
 colorTo: pink
 sdk: gradio
-sdk_version: 6.18.0
-python_version: '3.13'
 app_file: app.py
 pinned: false
+license: mit
+tags:
+  - build-small-hackathon
+  - openbmb
+  - voice-cloning
+  - tts
+  - indic
+  - thousand-token-wood
+  - multilingual
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🎙️ Polyglot Me
+
+Record ~10 seconds of your voice, type a line, and hear **yourself** say it in
+**English, Hindi, Telugu, and Tamil** — one cloned voice, every language.
+
+## Build approach
+
+Two models, zero configuration:
+
+1. **Sarvam-Translate** turns your English sentence into Hindi, Telugu, and Tamil
+   in a single API call.
+2. **VoxCPM2 (OpenBMB)** clones your voice from the 10-second reference clip and
+   generates each language independently. The voice is consistent across all four.
+
+All GPU work runs on **Modal** (serverless); this Space is CPU-only and calls Modal via
+`modal.Cls.from_name`. Output includes per-language audio players, a shareable waveform
+card, and a ready-to-post social caption.
+
+## Models (all ≤32B)
+
+| Role | Model | Params |
+|---|---|---|
+| Voice | [VoxCPM2](https://huggingface.co/openbmb/VoxCPM2) (OpenBMB) | — |
+| Praxy TTS base | [Praxel/praxy-voice-r6](https://huggingface.co/Praxel/praxy-voice-r6) | 544M |
+
+> ⏳ First run takes ~2 min while models warm up on Modal. Subsequent runs are fast.
+
+Built for the Hugging Face **Build Small** hackathon — *Thousand Token Wood* · OpenBMB.

app.py

@@ -0,0 +1,179 @@
+"""Polyglot Me — HF Space for App 2 (Thousand Token Wood).
+
+Record ~10s of your voice, type a line, hear yourself across English / Hindi /
+Telugu / Tamil. VoxCPM2 (OpenBMB) clones your voice for every language; Sarvam
+translates. All on Modal; this Space is CPU-only.
+
+Hackathon: Thousand Token Wood · OpenBMB · Modal.
+"""
+from __future__ import annotations
+import io
+import gradio as gr
+import matplotlib
+import modal
+import numpy as np
+import soundfile as sf
+from theme import build_css
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt  # noqa: E402
+
+MODAL_APP   = "praxy-voice"
+LANGS       = ["en", "hi", "te", "ta"]
+LANG_NAMES  = {"en": "English", "hi": "Hindi", "te": "Telugu", "ta": "Tamil"}
+LANG_SCRIPT = {"en": "Aa", "hi": "हि", "te": "తె", "ta": "த"}
+LANG_LABEL  = {"en": "LATIN", "hi": "DEVANAGARI", "te": "TELUGU", "ta": "TAMIL"}
+# Header colours: WCAG-AA verified with white text (≥5:1). Jewel tones on dark.
+LANG_HEADER = {"en": "#3B5998", "hi": "#C2410C", "te": "#0F766E", "ta": "#991B1B"}
+# Brighter accents for the matplotlib share-card waveforms (graphics, not text).
+LANG_WAVE   = {"en": "#8B7FF9", "hi": "#FF7A7A", "te": "#4FC878", "ta": "#5BB8F2"}
+
+_TRANSLATOR = modal.Cls.from_name(MODAL_APP, "SarvamTranslator")
+_VOX        = modal.Cls.from_name(MODAL_APP, "VoxCPM2Cloner")
+
+
+def _make_card(lines: dict, wav_paths: dict) -> str:
+    order = ["English", "Hindi", "Telugu", "Tamil"]
+    wave  = {LANG_NAMES[l]: LANG_WAVE[l] for l in LANGS}
+    fig = plt.figure(figsize=(10, 6.5), facecolor="#0E1020")
+    fig.text(.5, .965, "Polyglot Me", ha="center", va="top", color="white",
+             fontsize=23, fontweight="bold")
+    fig.text(.5, .905, "one voice · four languages", ha="center", va="top",
+             color="#8E8FB4", fontsize=12)
+    for i, lang in enumerate(order):
+        ax = fig.add_subplot(4, 1, i + 1); ax.set_facecolor("#0E1020")
+        c = wave[lang]; path = wav_paths.get(lang)
+        if path:
+            arr, _ = sf.read(path, dtype="float32")
+            xs = np.linspace(0, 1, len(arr))
+            ax.fill_between(xs, arr, alpha=.4, color=c); ax.plot(xs, arr, color=c, lw=.6, alpha=.9)
+        ax.set_xlim(0, 1)
+        ax.text(-.01, .5, lang, transform=ax.transAxes, ha="right", va="center",
+                color=c, fontsize=11, fontweight="bold")
+        snip = lines.get(lang, "")
+        if len(snip) > 64: snip = snip[:61] + "…"
+        ax.text(.015, .5, snip, transform=ax.transAxes, ha="left", va="center",
+                color="white", fontsize=9.5, alpha=.88)
+        ax.set_xticks([]); ax.set_yticks([])
+        for sp in ax.spines.values(): sp.set_visible(False)
+        ax.axvline(0, color=c, lw=4, solid_capstyle="round")
+    fig.subplots_adjust(left=.13, right=.98, top=.87, bottom=.03, hspace=.1)
+    out = "/tmp/polyglot_card.png"
+    fig.savefig(out, dpi=160, bbox_inches="tight", facecolor="#0E1020"); plt.close(fig)
+    return out
+
+
+def _lang_header_html(lang: str) -> str:
+    return (
+        f'<div style="background:{LANG_HEADER[lang]};padding:16px 20px;display:flex;'
+        f'align-items:center;gap:14px;border-radius:18px 18px 0 0;">'
+        f'<div style="font-family:\'Fraunces\',serif;font-size:32px;font-weight:700;line-height:1;'
+        f'color:rgba(255,255,255,.85);min-width:42px;text-align:center;">{LANG_SCRIPT[lang]}</div>'
+        f'<div style="display:flex;flex-direction:column;gap:3px;">'
+        f'<div style="font-size:14px;font-weight:800;color:#fff;letter-spacing:.02em;">{LANG_NAMES[lang]}</div>'
+        f'<div style="font-size:10px;font-weight:700;color:rgba(255,255,255,.82);'
+        f'letter-spacing:.1em;">{LANG_LABEL[lang]} SCRIPT</div></div></div>')
+
+
+def generate(ref_audio_path, line):
+    empty = [None, None, None, None, None, "Record a clip and type a line.", ""]
+    if not ref_audio_path or not (line and line.strip()):
+        return empty
+    with open(ref_audio_path, "rb") as f: ref_bytes = f.read()
+    translated = _TRANSLATOR().translate.remote(line, "en", ["hi", "te", "ta"])
+    lines = {"en": line, **translated}
+    outs = []
+    for lang in LANGS:
+        wb, _ = _VOX().clone.remote(text=lines[lang], ref_audio_bytes=ref_bytes)
+        path = f"/tmp/polyglot_{lang}.wav"
+        with open(path, "wb") as f: f.write(wb)
+        outs.append(path)
+    card = _make_card({LANG_NAMES[l]: lines[l] for l in LANGS},
+                      {LANG_NAMES[l]: outs[i] for i, l in enumerate(LANGS)})
+    transcript = "\n".join(f"{LANG_NAMES[l]}: {lines[l]}" for l in LANGS)
+    caption = ('I typed one line and heard myself say it in four languages 🎙️\n\n'
+               f'"{line}"\n\n' + transcript +
+               "\n\nBuilt with VoxCPM2 + Praxy for #BuildSmall #HuggingFace #PolyglotMe")
+    return outs + [card, transcript, caption]
+
+
+HERO = """
+<div style="border-radius:26px;padding:62px 28px 54px;text-align:center;position:relative;
+  overflow:hidden;min-height:300px;border:1px solid rgba(255,255,255,.08);
+  background:linear-gradient(135deg,#1a0050,#0d1a60,#003040,#0d1a00,#400010,#200040);
+  background-size:600% 600%;animation:gsh 12s ease infinite;">
+  <div style="position:absolute;top:-40px;left:-40px;width:210px;height:210px;border-radius:50%;
+    background:radial-gradient(circle,rgba(124,111,247,.4),transparent 70%);animation:orb 4s 0s ease-in-out infinite;"></div>
+  <div style="position:absolute;top:-20px;right:-20px;width:170px;height:170px;border-radius:50%;
+    background:radial-gradient(circle,rgba(247,111,111,.4),transparent 70%);animation:orb 4s .7s ease-in-out infinite;"></div>
+  <div style="position:absolute;bottom:-30px;left:30%;width:190px;height:190px;border-radius:50%;
+    background:radial-gradient(circle,rgba(79,200,120,.35),transparent 70%);animation:orb 4s 1.3s ease-in-out infinite;"></div>
+  <div style="position:absolute;bottom:-20px;right:10%;width:150px;height:150px;border-radius:50%;
+    background:radial-gradient(circle,rgba(79,184,247,.35),transparent 70%);animation:orb 4s 2s ease-in-out infinite;"></div>
+  <div style="position:relative;z-index:10;max-width:640px;margin:0 auto;">
+    <div style="font-size:60px;margin-bottom:10px;">🎙️</div>
+    <h1 style="font-family:'Fraunces',serif;font-size:clamp(2.6rem,7vw,4.4rem);font-weight:900;
+      color:#fff;margin:0 0 12px;letter-spacing:-.02em;line-height:1;
+      text-shadow:0 0 40px rgba(255,255,255,.3);">Polyglot Me</h1>
+    <p style="font-family:'Plus Jakarta Sans',sans-serif;font-size:clamp(1rem,2.5vw,1.2rem);
+      color:rgba(255,255,255,.82);margin:0 auto 18px;max-width:500px;line-height:1.6;">
+      Record ten seconds of your voice — hear <strong style="color:#fff;">yourself</strong>
+      speak English, Hindi, Telugu, and Tamil.</p>
+    <div style="display:flex;justify-content:center;gap:7px;flex-wrap:wrap;margin-bottom:12px;">
+      <span class="pf-pill" style="background:#3B5998;color:#fff;">ENGLISH</span>
+      <span class="pf-pill" style="background:#C2410C;color:#fff;">हिन्दी</span>
+      <span class="pf-pill" style="background:#0F766E;color:#fff;">తెలుగు</span>
+      <span class="pf-pill" style="background:#991B1B;color:#fff;">தமிழ்</span>
+    </div>
+    <div style="display:flex;justify-content:center;gap:8px;flex-wrap:wrap;">
+      <span class="pf-pill" style="background:rgba(255,255,255,.1);color:rgba(255,255,255,.8);border:1px solid rgba(255,255,255,.16);">VoxCPM2 · OPENBMB</span>
+      <span class="pf-pill" style="background:rgba(255,255,255,.1);color:rgba(255,255,255,.8);border:1px solid rgba(255,255,255,.16);">MODAL · SERVERLESS</span>
+    </div>
+    <p style="font-size:12.5px;color:rgba(255,255,255,.5);font-family:'Plus Jakarta Sans',sans-serif;margin:16px 0 0;">
+      ⏳ First run warms the model on Modal (~2–4 min). After that it's quick.</p>
+  </div>
+  <div style="position:absolute;bottom:0;left:0;right:0;height:3px;
+    background:linear-gradient(90deg,#3B5998,#C2410C,#0F766E,#991B1B);"></div>
+</div>
+"""
+
+EXTRA = """
+@keyframes gsh {0%{background-position:0% 50%}50%{background-position:100% 50%}100%{background-position:0% 50%}}
+@keyframes orb {0%,100%{transform:scale(1) translateY(0)}50%{transform:scale(1.06) translateY(-6px)}}
+.lang-card{padding:0!important;overflow:hidden!important;}
+#share-card img{border-radius:16px!important;border:1px solid rgba(255,255,255,.1)!important;}
+#caption-out textarea{border-left:3px solid #5BB8F2!important;font-size:13px!important;line-height:1.8!important;}
+#transcript-out textarea{font-size:13px!important;}
+.rainbow-hr{height:2px;background:linear-gradient(90deg,#3B5998,#C2410C,#0F766E,#991B1B);border:none;margin:18px 0;opacity:.5;border-radius:2px;}
+"""
+
+with gr.Blocks(title="Polyglot Me") as demo:
+    gr.HTML(f"<style>{build_css('polyglot', EXTRA)}</style>")
+    gr.HTML(HERO)
+
+    with gr.Row():
+        ref_in  = gr.Audio(sources=["microphone", "upload"], type="filepath",
+                           label="Your voice  (~10 seconds)", scale=1)
+        line_in = gr.Textbox(label="Say something  (in English)", lines=3, scale=2,
+                             placeholder="Good morning Amma, hope you slept well.")
+    speak_btn = gr.Button("🌍  Speak it in 4 languages", variant="primary", elem_id="cta")
+    gr.HTML('<hr class="rainbow-hr"/>')
+
+    audios = {}
+    with gr.Row():
+        for lang in LANGS:
+            with gr.Column(elem_classes=["lang-card"]):
+                gr.HTML(_lang_header_html(lang))
+                audios[lang] = gr.Audio(label="", type="filepath", show_label=False)
+
+    share_card = gr.Image(label="Share card", type="filepath", elem_id="share-card")
+    with gr.Row():
+        transcript_out = gr.Textbox(label="Translations", lines=4, elem_id="transcript-out", scale=1)
+        caption_out    = gr.Textbox(label="Caption  (copy for your post)", lines=6, elem_id="caption-out", scale=1)
+
+    speak_btn.click(generate, inputs=[ref_in, line_in],
+                    outputs=[audios["en"], audios["hi"], audios["te"], audios["ta"],
+                             share_card, transcript_out, caption_out])
+
+if __name__ == "__main__":
+    demo.launch()

praxy_router.py

@@ -0,0 +1,227 @@
+"""Praxy production inference router (v2 — adds IndicF5 + codemix).
+
+Single entry point that picks the right inference branch per language and
+codemix-detection. See `memory/project_indicf5_unblock_recipe_2026-04-27.md`
+and the v1-zero-shot scorecards in `evaluation/scorecards/indicf5_v1_*` for
+the empirical data motivating the routes below.
+
+Routing matrix (2026-04-27 v1):
+
+| Language     | Pure-script branch              | Codemix branch                  |
+|--------------|---------------------------------|---------------------------------|
+| Telugu (te)  | R6 LoRA (WER 0.034)             | translit → IndicF5 (WER 0.14)   |
+| Hindi (hi)   | vanilla Chatterbox + ref (0.025)| translit → IndicF5 (WER 0.20)   |
+| Tamil (ta)   | R6 LoRA (WER 0.041)             | translit → IndicF5 (WER 0.27)   |
+| English (en) | Chatterbox vanilla              | n/a                             |
+| Other Indic  | IndicF5 best-effort             | translit → IndicF5 best-effort  |
+
+`is_codemix(text)` returns True when ≥1 word is pure Latin (≥2 alphabetic
+chars). Single Latin chars or digits are not enough to trigger.
+
+The `translit → IndicF5` branch calls Haiku 4.5 to convert Latin English
+words into native-script phonetic spelling matching how Bollywood/Sarvam
+training data writes them ("WhatsApp" → "व्हाट्सऐप"), then sends the
+all-native-script string to IndicF5. This single fix dropped Hi codemix
+WER from 0.85 → 0.20 (76% relative drop) and Te codemix from 0.80 → 0.14
+(82% relative drop). See `serving/codemix_to_native_script.py`.
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+# Default reference audio per language. These are commercial-TTS-sourced
+# clips reused for the v1 demo; Pushpak will swap in Praxel-owned voices
+# (his + Ashwin's) before public launch.
+_DEFAULT_REFS: dict[str, str] = {
+    "te": "data/references/sarvam_te_female_9s.wav",
+    "ta": "data/references/sarvam_ta_male_11s.wav",
+    "hi": "data/references/sarvam_hi_female_10s.wav",
+    "en": "data/references/ashwin_10s.wav",  # Chatterbox base needs ref too
+}
+
+# Per-ref transcript cache (cleaned, no Whisper special tokens). IndicF5
+# requires ref audio + EXACT matching ref text.
+_DEFAULT_REF_TEXTS: dict[str, str] = {
+    "te": "మా తాతయ్య ప్రతి సాయంత్రం వరండాలో కూర్చుని తన చిన్నతనంలో జరిగిన కథలు చెబుతూ ఉంటారు మరియు మేము అందరూ కలిసి ఆసక్తిగా వినేవాళ్లం",
+    "ta": "எங்கள் தாத்தா தினமும் மாலையில் திண்ணையில் அமர்ந்து கொண்டு தன் சிறுவயதில் நடந்த கதைகளைச் சொல்லிக் கொண்டிருப்பார் நாங்கள் அனைவரும் சேர்ந்து ஆர்வமாகக் கேட்போம்",
+    "hi": "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",
+}
+
+# Per-language pure-script branch. "lora" = R6 LoRA (Chatterbox); "indicf5"
+# = ai4bharat/IndicF5 zero-shot; "vanilla" = base Chatterbox no-LoRA.
+_BRANCH: dict[str, str] = {
+    "te": "lora",       # R6 wins pure-Te (WER 0.034 < 0.08 IndicF5)
+    "ta": "lora",       # R6 wins pure-Ta (paper §V.1 0.041 LLM-WER)
+    "hi": "vanilla",    # vanilla Chatterbox + Cart-Hi ref + Config B = 0.025
+                        # WER (paper §V.1 ties Cartesia); IndicF5 zero-shot
+                        # is 0.13 — vanilla Chatterbox recipe wins on Hi.
+    "en": "vanilla",    # Chatterbox base is excellent at English
+}
+
+# Codemix always routes to IndicF5 with native-script preprocessing —
+# IndicF5's char-level tokenizer + the translit fix gives the only
+# working codemix recipe across our 8 architectural attempts.
+_CODEMIX_BRANCH = "indicf5_native"
+
+# Config B sampling overrides (Chatterbox-only — IndicF5 doesn't expose
+# these knobs). From TTS paper §5.2 sweep.
+CONFIG_B = dict(
+    exaggeration=0.7,
+    temperature=0.6,
+    min_p=0.1,
+    cfg_weight=0.5,
+    repetition_penalty=2.0,
+    top_p=1.0,
+)
+
+DEFAULT_R6_CKPT = "/cache/chatterbox_indic/round_6/step_8000.ckpt"
+
+# A "codemix" word is a run of ≥2 Latin alphabetic chars. Single letters
+# (acronym components) and digits don't trigger; numbers are normalised
+# elsewhere via Indic number expansion.
+_CODEMIX_WORD_RE = re.compile(r"[A-Za-z]{2,}")
+
+
+def is_codemix(text: str) -> bool:
+    """Returns True iff `text` contains at least one Latin word ≥2 chars
+    that needs transliteration before IndicF5 synthesis."""
+    return bool(_CODEMIX_WORD_RE.search(text))
+
+
+def _resolve_ref_audio(ref_audio_path: str | None, lang: str) -> str:
+    if ref_audio_path:
+        return ref_audio_path
+    fallback = _DEFAULT_REFS.get(lang)
+    if not fallback:
+        raise ValueError(f"No default reference voice for lang={lang!r}; supply ref_audio_path.")
+    return fallback
+
+
+def _resolve_ref_text(ref_text: str | None, lang: str) -> str:
+    if ref_text:
+        return ref_text
+    fallback = _DEFAULT_REF_TEXTS.get(lang, "")
+    return fallback
+
+
+def route(text: str, lang: str, ref_audio_path: str | None = None,
+          ref_text: str | None = None) -> dict:
+    """Return the inference-branch parameters for a given input.
+
+    Picks the route by combining (lang, codemix-detection). Returns a
+    dict with the branch identifier and the kwargs to forward to the
+    matching Modal entrypoint.
+    """
+    lang = lang.lower()
+    # Codemix routing applies only to Indic targets; English target with
+    # English text is not codemix.
+    cm = is_codemix(text) and lang != "en"
+    pure_branch = _BRANCH.get(lang, "indicf5")
+    branch = _CODEMIX_BRANCH if cm else pure_branch
+
+    ref_audio = _resolve_ref_audio(ref_audio_path, lang)
+    ref_t = _resolve_ref_text(ref_text, lang)
+
+    if branch == "lora":
+        return {
+            "branch": "lora",
+            "model": "chatterbox_r6_lora",
+            "ckpt_path": DEFAULT_R6_CKPT,
+            "use_bups": True,
+            "no_lora": False,
+            "ref_audio_path": ref_audio,
+            "normalize_numbers": True,
+            "language_code": lang,
+            **CONFIG_B,
+        }
+    if branch == "vanilla":
+        return {
+            "branch": "vanilla",
+            "model": "chatterbox_base",
+            "ckpt_path": DEFAULT_R6_CKPT,  # unused with no_lora
+            "use_bups": False,
+            "no_lora": True,
+            "ref_audio_path": ref_audio,
+            "normalize_numbers": True,
+            "language_code": lang,
+            **CONFIG_B,
+        }
+    if branch == "indicf5":
+        return {
+            "branch": "indicf5",
+            "model": "indicf5_zeroshot",
+            "ref_audio_path": ref_audio,
+            "ref_text": ref_t,
+            "language_code": lang,
+        }
+    if branch == "indicf5_native":
+        # Caller should run text through `serving.codemix_to_native_script
+        # .transliterate_codemix(text, lang)` before invoking the synth.
+        return {
+            "branch": "indicf5_native",
+            "model": "indicf5_native_codemix",
+            "ref_audio_path": ref_audio,
+            "ref_text": ref_t,
+            "language_code": lang,
+            "preprocess": "transliterate_codemix_to_native",
+        }
+    raise ValueError(f"Unknown branch {branch!r}")
+
+
+def synthesize(text: str, lang: str, ref_audio_path: str | None = None,
+               ref_text: str | None = None) -> tuple[bytes, int]:
+    """Production-grade single-utterance synthesis. Wraps `route()` and
+    runs the chosen Modal entrypoint. Returns (wav_bytes, sample_rate).
+
+    Note: this is a thin convenience wrapper. For batch eval/training
+    use the underlying Modal entrypoints directly to avoid per-call
+    Modal overhead.
+    """
+    plan = route(text, lang, ref_audio_path=ref_audio_path, ref_text=ref_text)
+    branch = plan["branch"]
+
+    if branch == "indicf5_native":
+        from serving.codemix_to_native_script import transliterate_codemix
+        text = transliterate_codemix(text, lang)
+
+    if branch in ("indicf5", "indicf5_native"):
+        from serving.modal_app import IndicF5TTS
+        synth = IndicF5TTS()
+        from pathlib import Path as _P
+        ref_bytes = _P(plan["ref_audio_path"]).read_bytes()
+        return synth.synthesize.remote(
+            text=text,
+            ref_audio_bytes=ref_bytes,
+            ref_text=plan["ref_text"],
+        )
+
+    # lora / vanilla — Chatterbox path (existing implementation).
+    from serving.modal_app import PraxyChatterboxLoRA
+    import modal as _modal
+    env = {
+        "PRAXY_CKPT_PATH": plan["ckpt_path"],
+        "PRAXY_USE_BUPS": "1" if plan["use_bups"] else "0",
+        "PRAXY_NO_LORA": "1" if plan["no_lora"] else "0",
+    }
+    synth = PraxyChatterboxLoRA.with_options(secrets=[_modal.Secret.from_dict(env)])()
+    from pathlib import Path as _P
+    ref_bytes = _P(plan["ref_audio_path"]).read_bytes() if plan["ref_audio_path"] else None
+    if plan.get("normalize_numbers") and lang in {"te", "ta", "hi", "bn", "gu", "mr", "kn", "ml"}:
+        from praxy.linguistics.indic_numbers import normalize_indic_text
+        text = normalize_indic_text(text, lang)
+    return synth.synthesize.remote(
+        text=text,
+        language_code=lang,
+        ref_audio_bytes=ref_bytes,
+        exaggeration=CONFIG_B["exaggeration"],
+        cfg_weight=CONFIG_B["cfg_weight"],
+        temperature=CONFIG_B["temperature"],
+        repetition_penalty=CONFIG_B["repetition_penalty"],
+        min_p=CONFIG_B["min_p"],
+        top_p=CONFIG_B["top_p"],
+    )
+
+
+__all__ = ["route", "synthesize", "is_codemix", "CONFIG_B", "DEFAULT_R6_CKPT"]

requirements.txt

@@ -0,0 +1,5 @@
+gradio>=4.40
+modal
+numpy
+soundfile
+matplotlib

theme.py

@@ -0,0 +1,163 @@
+"""Praxy Field Guide — shared design system for the Build-Small hackathon apps.
+
+One CSS framework, themed per app via tokens. Injected via gr.HTML('<style>…</style>')
+because Gradio 6.0 dropped the css= argument from gr.Blocks(). All colour pairings
+are chosen to pass WCAG AA (body text >=4.5:1, large text / UI >=3:1).
+
+A copy of this file lives in each hf_space_* directory (the spaces upload
+independently and cannot share a module). Keep them in sync.
+"""
+from __future__ import annotations
+
+# Per-app design tokens. Each value is a CSS colour; contrast verified vs bg/accent.
+THEMES = {
+    # Nidra — night sky. Dark is thematically right; text is solid (not low-alpha).
+    "nidra": {
+        "bg": "#0B1026", "bg2": "#070A1C", "surface": "#161C3F", "surface2": "#1E2550",
+        "ink": "#F4EFDE", "muted": "#AEB4D6", "faint": "#8086AE",
+        "accent": "#F5C842", "accent_ink": "#1A1304", "accent2": "#9E86FF",
+        "border": "rgba(245,200,66,.20)", "ring": "rgba(245,200,66,.45)",
+        "display": "Fraunces", "body": "Plus Jakarta Sans",
+    },
+    # Read-it-to-Amma — warm morning parchment. Accessibility app -> high contrast.
+    "readit": {
+        "bg": "#FBF3E3", "bg2": "#F6E9CE", "surface": "#FFFCF5", "surface2": "#F5EDE0",
+        "ink": "#2D1306", "muted": "#6B3D1E", "faint": "#9A7550",
+        "accent": "#C2410C", "accent_ink": "#FFFFFF", "accent2": "#7B1F2E",
+        "border": "rgba(120,60,0,.18)", "ring": "rgba(194,65,12,.35)",
+        "display": "Fraunces", "body": "Plus Jakarta Sans",
+    },
+    # Polyglot Me — festival; refined dark with four AA language accents.
+    "polyglot": {
+        "bg": "#0E1020", "bg2": "#080A18", "surface": "#181B33", "surface2": "#20243F",
+        "ink": "#F1EEF8", "muted": "#AFB0CC", "faint": "#7E80A2",
+        "accent": "#8B7FF9", "accent_ink": "#0B0B1A", "accent2": "#5BB8F2",
+        "border": "rgba(255,255,255,.10)", "ring": "rgba(139,127,249,.45)",
+        "display": "Fraunces", "body": "Plus Jakarta Sans",
+    },
+    # Parliament of Owls — naturalist field guide; parchment + forest + brass.
+    # Palette WCAG-verified: ink/bg 16.6:1, muted/bg 7.3:1, forest-accent/cream 7+:1.
+    "owls": {
+        "bg": "#F0E6CC", "bg2": "#E6D8B6", "surface": "#FAF5EB", "surface2": "#F2E9D2",
+        "ink": "#221E14", "muted": "#5C4A1E", "faint": "#8A7B5C",
+        "accent": "#2F5D44", "accent_ink": "#F7F1E1", "accent2": "#8B6914",
+        "border": "rgba(60,50,30,.22)", "ring": "rgba(47,93,68,.40)",
+        "display": "Fraunces", "body": "Plus Jakarta Sans",
+    },
+}
+
+_FONTS = ("https://fonts.googleapis.com/css2?"
+          "family=Fraunces:ital,opsz,wght@0,9..144,400;0,9..144,600;0,9..144,700;0,9..144,900;"
+          "1,9..144,400;1,9..144,600&"
+          "family=Plus+Jakarta+Sans:wght@400;500;600;700;800&"
+          "family=Noto+Sans:wght@400;500;600&display=swap")
+
+
+def build_css(theme: str, extra: str = "") -> str:
+    """Return the full themed stylesheet for `theme` (a key in THEMES)."""
+    t = THEMES[theme]
+    return f"""
+@import url('{_FONTS}');
+
+:root {{
+  --bg:{t['bg']}; --bg2:{t['bg2']}; --surface:{t['surface']}; --surface2:{t['surface2']};
+  --ink:{t['ink']}; --muted:{t['muted']}; --faint:{t['faint']};
+  --accent:{t['accent']}; --accent-ink:{t['accent_ink']}; --accent2:{t['accent2']};
+  --border:{t['border']}; --ring:{t['ring']};
+  --display:'{t['display']}',Georgia,serif; --body:'{t['body']}','Noto Sans',sans-serif;
+  --radius:18px; --gap:18px;
+}}
+
+/* ---- fill the whole viewport, widen the column ---- */
+html, body, gradio-app, .gradio-container, .app, #root {{
+  background: var(--bg) !important;
+}}
+body, .gradio-container {{ font-family: var(--body) !important; color: var(--ink) !important; }}
+.gradio-container {{
+  max-width: 1180px !important; margin: 0 auto !important;
+  padding-left: 20px !important; padding-right: 20px !important;
+}}
+.gradio-container, .gradio-container * {{ box-sizing: border-box; }}
+
+/* headings + prose */
+h1,h2,h3 {{ font-family: var(--display) !important; color: var(--ink) !important; }}
+
+/* ---- surfaces / cards ---- */
+.gr-form, .gr-block, .gr-panel, .block, .contain, .gr-box, .form {{
+  background: var(--surface) !important;
+  border: 1px solid var(--border) !important;
+  border-radius: var(--radius) !important;
+}}
+.gap, .panel, .wrap {{ gap: var(--gap) !important; }}
+
+/* ---- labels ---- */
+label > span, .label-wrap > span, span[data-testid="block-info"] {{
+  color: var(--muted) !important; font-size: 11px !important;
+  font-weight: 700 !important; letter-spacing: .09em !important;
+  text-transform: uppercase !important; font-family: var(--body) !important;
+}}
+
+/* ---- inputs ---- */
+input[type=text], input[type=number], textarea, select, .gr-text-input {{
+  background: var(--surface2) !important;
+  border: 1.5px solid var(--border) !important;
+  border-radius: 12px !important;
+  color: var(--ink) !important; font-size: 15px !important;
+  font-family: var(--body) !important;
+}}
+input::placeholder, textarea::placeholder {{ color: var(--faint) !important; opacity: 1; }}
+input:focus, textarea:focus, select:focus {{
+  border-color: var(--accent) !important; outline: none !important;
+  box-shadow: 0 0 0 3px var(--ring) !important;
+}}
+input[type=range] {{ accent-color: var(--accent) !important; }}
+input[type=checkbox] {{ accent-color: var(--accent) !important; width:18px; height:18px; }}
+
+/* ---- accessible focus ring for keyboard users ---- */
+:focus-visible {{ outline: 3px solid var(--ring) !important; outline-offset: 2px !important; }}
+
+/* ---- primary CTA ---- */
+button.primary, .gr-button-primary, #cta button {{
+  background: var(--accent) !important; color: var(--accent-ink) !important;
+  font-weight: 800 !important; font-size: 16px !important; letter-spacing: .02em !important;
+  font-family: var(--body) !important; border: none !important;
+  border-radius: 14px !important; padding: 16px 30px !important; width: 100% !important;
+  cursor: pointer !important;
+  box-shadow: 0 6px 22px rgba(0,0,0,.18) !important;
+  transition: transform .15s ease, box-shadow .25s ease, filter .2s ease !important;
+}}
+button.primary:hover, .gr-button-primary:hover, #cta button:hover {{
+  transform: translateY(-2px) !important; filter: brightness(1.05) !important;
+  box-shadow: 0 10px 34px rgba(0,0,0,.28) !important;
+}}
+#cta button:active {{ transform: translateY(0) !important; }}
+
+/* secondary buttons */
+.gr-button-secondary {{
+  background: var(--surface2) !important; color: var(--ink) !important;
+  border: 1.5px solid var(--border) !important; border-radius: 12px !important;
+  font-family: var(--body) !important; font-weight: 600 !important;
+}}
+
+/* audio + image blocks */
+.waveform-container, .gr-audio, audio {{ border-radius: 12px !important; }}
+
+/* tidy chrome */
+footer {{ display: none !important; }}
+.gradio-container .prose a {{ color: var(--accent2) !important; }}
+::-webkit-scrollbar {{ width: 8px; height: 8px; }}
+::-webkit-scrollbar-track {{ background: var(--bg2); }}
+::-webkit-scrollbar-thumb {{ background: var(--accent); border-radius: 6px; opacity:.6; }}
+
+/* shared pill / divider helpers used inside gr.HTML output */
+.pf-pill {{
+  display:inline-block; padding:5px 14px; border-radius:20px; font-size:11px;
+  font-weight:700; letter-spacing:.06em; font-family:var(--body);
+}}
+.pf-divider {{ height:1px; background:linear-gradient(90deg,transparent,var(--border),transparent); margin:18px 0; border:none; }}
+
+@media (max-width: 760px) {{
+  .gradio-container {{ padding-left: 12px !important; padding-right: 12px !important; }}
+}}
+{extra}
+"""