README.md

---
title: GEPARD
emoji: 🐆
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.20.0
python_version: '3.12'
app_file: app.py
pinned: false
license: apache-2.0
short_description: Spotted text-to-speech with voice cloning and text-CFG
---

# 🐆 GEPARD TTS

Inference Space for the **Gepard** autoregressive speech model
(`nineninesix/gepard-1.0` — a Qwen3.5 backbone with 32 FSQ audio
heads, a Q-Former voice-cloning compressor, and DPO post-training).

## Features

- **Preset speakers** — bundled voices stored as pre-encoded codec tokens
  (`speakers/*.pt`), no codec pass needed at request time.
- **Voice cloning** — record with the microphone or upload a clip; the
  reference is encoded with the NeMo nano codec and compressed into 8 speaker
  prefix tokens.
- **Text classifier-free guidance** — `cfg_scale`/`cfg_frames` sharpen text
  and voice adherence (defaults: `temperature=0.3`, `cfg_scale=3`).
- All generation knobs are exposed under **Generation settings**.

## Architecture

The Space runs on `gradio.Server` — a FastAPI app with Gradio's queueing
engine on top. The custom dark-themed UI in `index.html` is served at `/`
via `@app.get("/")`; the synthesis pipeline is exposed at
`@app.api("/synthesize")`, so requests flow through Gradio's queue
(concurrency control, ZeroGPU allocation, `gradio_client` compatibility)
while the frontend stays a self-contained HTML/CSS/JS bundle.

## Configuration

`config.yaml` selects the checkpoint, codec, preset speakers and default
generation parameters — re-point the Space without touching code.

## Notes

- The model repo is private: set the `HF_TOKEN` Space secret.
- ZeroGPU: the model is loaded once at startup; each request only runs
  generation inside the GPU context.
- `create_env.py` orchestrates a 4-step install at app startup — it must
  run before any ML import in `app.py`:
  1. Pin `huggingface-hub>=1.2,<2.0` (compatible with both gradio 6.20
     and `nemo-toolkit`).
  2. `pip install --no-deps nemo-toolkit[tts]==2.4.0` (avoiding a downgrade
     of hub by the resolver).
  3. Force-reinstall `transformers==5.3.0` (Qwen3.5 backbone that the
     Gepard checkpoint was trained on).
  4. Cap `numpy<2.0` so the codec/NeMo stack stays on numpy 1.x.
- `nemo-toolkit` is installed at runtime (not in `requirements.txt`) to
  keep gradio 6.20's `huggingface-hub>=1.2` constraint resolvable at
  build time — NeMo's `transformers<=4.52` would otherwise pull hub<1.0.