README.md

---
title: Kokoro TTS
emoji: 🗣️
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.6.0
app_file: app.py
pinned: false
python_version: "3.11"
---

# Kokoro TTS (and optional LFM3 for comparison)

Kokoro-82M answers on `/` (Gradio) and `POST /tts/stream` (raw PCM16 mono, streamed).

## Optional second engine: LFM3-Audio-2B

Off by default. To compare it with Kokoro in the same Space, set the Space variable
`LFM3_ENABLED=1`. It adds routes with the same contract under `/lfm3/tts`:

- `POST /lfm3/tts/stream`: `{"text", "voice": "female" | "male" | "af_heart"-style, "sample_rate": 16000 | 24000}`;
  `speed` is ignored (LFM3 has no speed control).
- `POST /lfm3/tts/bench`: seconds spent synthesizing vs seconds of audio, per piece.
  The first call loads the model, so read the second.
- `GET /lfm3/tts/health`.

The ~8 GB model (float32 on CPU) loads on the first LFM3 request, not at start-up, so Kokoro
is unaffected until then. LFM3 and Kokoro share the 2 vCPUs: benchmark them one at a time.
`TTS_API_KEY`, when set, protects both engines' POST routes.

To make the tutor use LFM3, point `KOKORO_HF_SPACE_URL` at `https://<space>.hf.space/lfm3`
(the client appends `/tts/stream`). Compare both with
`python -m scripts.compare_tts_spaces --space kokoro=https://<space>.hf.space --space lfm3=https://<space>.hf.space/lfm3`.

License of the model: LFM Open License v1.0 (commercial use only below US$10M annual revenue).