echoyard / README.md
loay's picture
Update README.md
881d089 verified
|
Raw
History Blame Contribute Delete
2.27 kB
---
title: EchoYard
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
python_version: "3.10"
license: apache-2.0
tags:
- build-small
- gradio
- backyard-ai
- best-minicpm-build
- tiny-titan
- openbmb
- minicpm
- voxcpm2
- text-to-speech
- language-learning
- speech-practice
- track:backyard
- sponsor:openbmb
- achievement:offgrid
---
# EchoYard
EchoYard is a tiny listen-and-repeat speaking practice app. Pick a language, level, and voice style; listen to a short reference voice; record yourself; then get a simple score and practical next step.
It is built for Build Small as a practical daily speaking tool: short voice prompts, quick browser recording, and friendly feedback powered by compact open models.
## Demo Video
[![Watch the EchoYard demo](https://img.youtube.com/vi/1K7IabUxXvw/maxresdefault.jpg)](https://youtu.be/1K7IabUxXvw)
[Watch the demo on YouTube](https://youtu.be/1K7IabUxXvw)
## Community Post
[Read the EchoYard community post on Hugging Face](https://huggingface.co/posts/loay/338571827537212)
## How It Works
1. Choose a supported language, level, and voice style.
2. Leave the text blank for a short starter line, or type a custom sentence.
3. Press Make voice to load a prebuilt sample or synthesize a short reference.
4. Record your attempt in the browser.
5. Press Score to compare the attempt with the reference.
6. Read the short feedback and try the next line.
## Build Notes
- `openbmb/VoxCPM2` generates multilingual reference speech for custom lines and the prebuilt starter library.
- `openbmb/MiniCPM5-1B` returns the final score, sub-scores, short feedback, and next practice line.
- `facebook/wav2vec2-base-960h` adds English-only transcript and word-match evidence.
- Deterministic audio features provide duration, active speech, timing, rhythm, pitch, and sound-shape evidence.
- A custom browser recorder captures microphone audio and the server converts it to scoring-ready WAV.
- A prebuilt reference-audio dataset keeps the first starter example fast for all supported languages and voice styles.
## Runtime And Privacy
Generated custom audio is stored only in the active Space runtime filesystem. User recordings are processed for scoring and are not written to a database.