Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,6 @@ 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
+audio/featured_268bc1833e83.wav filter=lfs diff=lfs merge=lfs -text
+audio/featured_46c861c190e2.wav filter=lfs diff=lfs merge=lfs -text
+audio/featured_df287dfc1406.wav filter=lfs diff=lfs merge=lfs -text

DEMO.md

@@ -0,0 +1,30 @@
+# Say It Back — recording + submission kit
+
+**Space:** https://huggingface.co/spaces/Tonykip/say-it-back
+
+## Before you hit record (speed)
+1. Open the Space, let it fully load.
+2. Click **2–3 featured phrases once** to warm everything (they're instant after).
+3. Type one live phrase and run it once to warm the translate endpoint (~20s first time, fast after — it caches).
+Then record. On camera everything will feel snappy.
+
+## Demo video (~60–90s)
+1. **The why (10s):** "Kalenjin is fading in my family — my wife understands a little but thinks in English. I built a tool so she can just say a word in English and hear how to say it in Kalenjin."
+2. **Instant phrases (15s):** click **Thank you → _Kongoi_**, **Water → _Peek_**, **I love you → _achamin_** — each plays back in the Cheps voice immediately.
+3. **The live moment (25s):** click the mic, say **"Good morning, how are you?"** → it transcribes → shows **_Amunee karon?_** → speaks it back in Kalenjin. That's the wow.
+4. **The human close (15s):** your wife (or you) using it for real — that footage IS the Backyard-track story.
+
+## Social post draft (edit freely)
+> Kalenjin is slipping away between generations in my own family. So this weekend I built **Say It Back** 🗣️ — say a word in English, hear it spoken back in Kalenjin, in a fine-tuned Kalenjin voice.
+>
+> Whisper (ASR) → EN→Kalenjin translation → Cheps TTS, wired up in Gradio.
+> Built for #BuildSmall (@huggingface × @Gradio). Models on @modal. 🇰🇪
+>
+> Try it: https://huggingface.co/spaces/Tonykip/say-it-back
+
+## Submission checklist
+- [ ] Gradio Space public + working ✅ (this Space)
+- [ ] Demo video recorded
+- [ ] Social post published (tag HF, Gradio, Modal)
+- [ ] Submit the Space link via the hackathon form / channel
+- [ ] (Optional) ask org admins to add you as writer on `build-small-hackathon` if org-hosting is required, then duplicate the Space there.

README.md

@@ -1,13 +1,41 @@
 ---
-title: Say It Back
-emoji: 📊
-colorFrom: pink
-colorTo: yellow
+title: Say It Back in Kalenjin
+emoji: 🗣️
+colorFrom: yellow
+colorTo: red
 sdk: gradio
-sdk_version: 6.18.0
-python_version: '3.13'
+sdk_version: 5.49.1
 app_file: app.py
 pinned: false
+short_description: Say a word in English, hear it in Kalenjin
+tags:
+  - build-small-hackathon
+  - backyard-ai
+  - tiny-titan
+  - best-demo
+  - speech
+  - low-resource-languages
+  - kalenjin
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🗣️ Say It Back in Kalenjin
+
+Say a word in English. Hear it in Kalenjin, in a real Kalenjin voice.
+
+Kalenjin is slipping away in my family. My wife thinks in English and wants to
+learn it; this lets her say a word and hear how to say it back in Kalenjin.
+
+## How it works
+Speak (or type) English → Whisper transcribes it → an English→Swahili→Kalenjin
+cascade translates it → **Cheps** (a Kalenjin TTS I fine-tuned) speaks it.
+Common phrases are pre-rendered so they play instantly.
+
+## Build Small (Hugging Face × Gradio)
+- **Track:** 🏡 Backyard AI.
+- **🐣 Tiny Titan:** every model is ≤4B — Whisper 74M (ASR), NLLB 600M ×2 in the
+  cascade (translation), Orpheus 3B + LoRA (the Cheps voice). Small weights, real
+  impact for a language with almost no tooling.
+- **Demo video:** _TODO: link_
+- **Social post:** _TODO: link_
+
+Models run on Modal; this Gradio Space ties them together.

app.py

@@ -0,0 +1,173 @@
+"""Walutik (Say It Back) — say a word in English, hear it in Kalenjin.
+
+English speech --(Whisper)--> English text
+              --(Kalenjin cascade, Modal)--> Kalenjin text
+              --(Cheps TTS, Modal)--> Kalenjin audio (auto-plays)
+
+Built for the HF × Gradio "Build Small" hackathon — Backyard AI track.
+"""
+
+from __future__ import annotations
+
+import functools
+import hashlib
+import json
+import os
+
+import gradio as gr
+import requests
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+
+TRANSLATE_URL = os.environ.get(
+    "MODAL_TRANSLATE_URL",
+    "https://tonykipkemboi--kalenjin-cascade-api-textcascade-translate.modal.run",
+)
+TTS_URL = os.environ.get(
+    "MODAL_SYNTHESIZE_URL",
+    "https://tonykipkemboi--kalenjin-tts-serve-kalenjintts-fastapi-app.modal.run/synthesize",
+)
+INFERENCE_TOKEN = os.environ.get("INFERENCE_TOKEN", "")  # Space secret; never committed
+
+AUDIO_DIR = os.path.join(HERE, "audio")
+os.makedirs(AUDIO_DIR, exist_ok=True)
+
+try:
+    with open(os.path.join(HERE, "phrasebook.json"), encoding="utf-8") as f:
+        PHRASEBOOK = json.load(f)
+except FileNotFoundError:
+    PHRASEBOOK = {}
+FEATURED = list(PHRASEBOOK.keys()) or [
+    "Thank you", "Good morning", "How are you?", "I love you",
+    "Water", "Milk", "My child", "Come and eat",
+]
+
+# ── English ASR (small Whisper, loaded once) ─────────────────────────────────
+_whisper = None
+
+
+def _get_whisper():
+    global _whisper
+    if _whisper is None:
+        from faster_whisper import WhisperModel
+
+        _whisper = WhisperModel("base.en", device="cpu", compute_type="int8")
+    return _whisper
+
+
+def transcribe_english(audio_path):
+    if not audio_path:
+        return ""
+    segments, _ = _get_whisper().transcribe(audio_path, language="en", beam_size=1)
+    return " ".join(s.text for s in segments).strip()
+
+
+# ── translate + synthesize (cached for instant repeats) ──────────────────────
+@functools.lru_cache(maxsize=512)
+def translate_en_to_kal(text):
+    r = requests.post(TRANSLATE_URL, json={"text": text}, timeout=70)
+    r.raise_for_status()
+    d = r.json()
+    return (d.get("kalenjin", "").strip(), d.get("swahili", "").strip())
+
+
+@functools.lru_cache(maxsize=512)
+def synthesize_kalenjin(text):
+    headers = {"Content-Type": "application/json"}
+    if INFERENCE_TOKEN:
+        headers["Authorization"] = f"Bearer {INFERENCE_TOKEN}"
+    r = requests.post(TTS_URL, json={"text": text}, headers=headers, timeout=150)
+    r.raise_for_status()
+    key = hashlib.sha256(text.encode()).hexdigest()[:16]
+    path = os.path.join(AUDIO_DIR, f"tts_{key}.wav")
+    with open(path, "wb") as f:
+        f.write(r.content)
+    return path
+
+
+def say_it_back(typed_text, audio_path):
+    english = (typed_text or "").strip()
+    if not english and audio_path:
+        yield "listening…", "", "", None
+        english = transcribe_english(audio_path)
+    if not english:
+        yield "Say or type something in English.", "", "", None
+        return
+
+    if english in PHRASEBOOK:
+        e = PHRASEBOOK[english]
+        yield english, e["kalenjin"], f"Swahili: {e['swahili']}", os.path.join(HERE, e["wav"])
+        return
+
+    yield english, "translating…", "", None
+    try:
+        kalenjin, swahili = translate_en_to_kal(english)
+    except Exception as e:  # noqa: BLE001
+        yield english, f"translation failed: {e}", "", None
+        return
+    if not kalenjin:
+        yield english, "no translation came back.", "", None
+        return
+
+    yield english, kalenjin, f"Swahili: {swahili}", None
+    try:
+        wav = synthesize_kalenjin(kalenjin)
+    except Exception as e:  # noqa: BLE001
+        yield english, kalenjin, f"Swahili: {swahili} · voice failed: {e}", None
+        return
+    yield english, kalenjin, f"Swahili: {swahili}", wav
+
+
+def featured_click(phrase):
+    e = PHRASEBOOK.get(phrase)
+    if e:
+        return phrase, e["kalenjin"], f"Swahili: {e['swahili']}", os.path.join(HERE, e["wav"])
+    kalenjin, swahili = translate_en_to_kal(phrase)
+    return phrase, kalenjin, f"Swahili: {swahili}", synthesize_kalenjin(kalenjin)
+
+
+# ── UI (mobile-first single column) ──────────────────────────────────────────
+CSS = """
+.gradio-container { max-width: 640px !important; margin: auto !important; }
+.kal-out textarea { font-size: 32px !important; font-weight: 700 !important; color:#b34700 !important; text-align:center; }
+.en-out textarea { font-size: 18px !important; }
+footer { visibility: hidden; }
+"""
+
+with gr.Blocks(title="Say It Back in Kalenjin", theme=gr.themes.Soft(primary_hue="orange"), css=CSS) as demo:
+    gr.Markdown(
+        "# 🗣️ Say It Back in Kalenjin\n"
+        "**Say a word in English. Hear it in Kalenjin.**\n\n"
+        "Kalenjin is slipping away in my family, so I built this to help us learn "
+        "it back. The voice is Cheps, a Kalenjin TTS I fine-tuned."
+    )
+
+    with gr.Group():
+        mic = gr.Audio(sources=["microphone"], type="filepath", label="🎤 Say it in English")
+        txt = gr.Textbox(label="…or type English", placeholder="e.g. good morning, how are you?")
+        go = gr.Button("Say it back →", variant="primary", size="lg")
+
+    with gr.Group():
+        kal_out = gr.Textbox(label="Kalenjin", interactive=False, elem_classes="kal-out")
+        audio_out = gr.Audio(label="🔊 Kalenjin voice (Cheps)", autoplay=True, interactive=False)
+        en_out = gr.Textbox(label="You said", interactive=False, elem_classes="en-out")
+        via = gr.Markdown("")
+
+    gr.Markdown("**Tap a phrase to hear it:**")
+    with gr.Row():
+        for _phrase in FEATURED:
+            gr.Button(_phrase, size="sm").click(
+                featured_click, inputs=gr.State(_phrase),
+                outputs=[en_out, kal_out, via, audio_out],
+            )
+
+    gr.Markdown(
+        "<sub>Tiny by design — every model is small: Whisper 74M (speech), "
+        "NLLB 600M (translation), Cheps / Orpheus 3B (voice). All under 4B.</sub>"
+    )
+
+    go.click(say_it_back, inputs=[txt, mic], outputs=[en_out, kal_out, via, audio_out])
+    txt.submit(say_it_back, inputs=[txt, mic], outputs=[en_out, kal_out, via, audio_out])
+
+if __name__ == "__main__":
+    demo.queue().launch()

audio/featured_268bc1833e83.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b82e864c0cd618cc33f00e82581bb203709bee54ece8398ca97856b9c5c52002
+size 121596

audio/featured_46c861c190e2.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af4c9018e27b416b1d898ca960617a3f31dfd2d7c15bbc57dea926b63b81cd0d
+size 155950

audio/featured_df287dfc1406.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6518e24cf758b481c9c8e670bc9ac55a1394367f0b75bd4d3dc349f399c3ede3
+size 105684

phrasebook.json

@@ -0,0 +1,42 @@
+{
+  "Thank you": {
+    "kalenjin": "Kongoi",
+    "swahili": "Asanteni",
+    "wav": "audio/featured_268bc1833e83.wav"
+  },
+  "Good morning": {
+    "kalenjin": "Amunee karon",
+    "swahili": "Habari za asubuhi",
+    "wav": "audio/featured_90a90a48e23d.wav"
+  },
+  "How are you?": {
+    "kalenjin": "Inye itestai ano?",
+    "swahili": "Wewe unaendeleaje?",
+    "wav": "audio/featured_df287dfc1406.wav"
+  },
+  "I love you": {
+    "kalenjin": "achamin",
+    "swahili": "Nakupenda",
+    "wav": "audio/featured_c33084feaa65.wav"
+  },
+  "Water": {
+    "kalenjin": "Peek",
+    "swahili": "Maji",
+    "wav": "audio/featured_7ca7dea90680.wav"
+  },
+  "Milk": {
+    "kalenjin": "Chego",
+    "swahili": "Maziwa",
+    "wav": "audio/featured_8189b8409b18.wav"
+  },
+  "My child": {
+    "kalenjin": "Lakwenyu",
+    "swahili": "Mtoto wangu",
+    "wav": "audio/featured_ea4c0f6bd91d.wav"
+  },
+  "Come and eat": {
+    "kalenjin": "Nyon any amisye",
+    "swahili": "Njoo na kula",
+    "wav": "audio/featured_46c861c190e2.wav"
+  }
+}

prerender.py

@@ -0,0 +1,44 @@
+"""Pre-render the featured phrases (English -> Kalenjin -> Cheps WAV) and write
+a phrasebook.json + audio/featured_*.wav bundle so they play back INSTANTLY in
+the app (no live Modal call on the hot demo path). Run locally before deploy:
+
+    INFERENCE_TOKEN=... python prerender.py
+"""
+
+import hashlib
+import json
+import os
+
+import requests
+
+TRANSLATE_URL = "https://tonykipkemboi--kalenjin-cascade-api-textcascade-translate.modal.run"
+TTS_URL = "https://tonykipkemboi--kalenjin-tts-serve-kalenjintts-fastapi-app.modal.run/synthesize"
+TOKEN = os.environ["INFERENCE_TOKEN"]
+
+FEATURED = [
+    "Thank you", "Good morning", "How are you?", "I love you",
+    "Water", "Milk", "My child", "Come and eat",
+]
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+AUDIO = os.path.join(HERE, "audio")
+os.makedirs(AUDIO, exist_ok=True)
+
+book = {}
+for phrase in FEATURED:
+    t = requests.post(TRANSLATE_URL, json={"text": phrase}, timeout=70).json()
+    kal, sw = t.get("kalenjin", "").strip(), t.get("swahili", "").strip()
+    wav = requests.post(
+        TTS_URL, json={"text": kal},
+        headers={"Authorization": f"Bearer {TOKEN}"}, timeout=150,
+    ).content
+    key = hashlib.sha256(phrase.encode()).hexdigest()[:12]
+    fname = f"featured_{key}.wav"
+    with open(os.path.join(AUDIO, fname), "wb") as f:
+        f.write(wav)
+    book[phrase] = {"kalenjin": kal, "swahili": sw, "wav": f"audio/{fname}"}
+    print(f"  {phrase!r:22} -> {kal!r:28} ({len(wav)} bytes)")
+
+with open(os.path.join(HERE, "phrasebook.json"), "w") as f:
+    json.dump(book, f, ensure_ascii=False, indent=2)
+print(f"wrote phrasebook.json ({len(book)} phrases)")

requirements.txt

@@ -0,0 +1,2 @@
+requests>=2.31
+faster-whisper>=1.1.0