Prepare encoder release documentation
Browse files
README.md
CHANGED
|
@@ -13,32 +13,31 @@ with [HeartCodec-oss-20260123](https://huggingface.co/HeartMuLa/HeartCodec-oss-2
|
|
| 13 |
which supplies the shared RVQ quantizer and audio decoder. Together they support
|
| 14 |
audio → eight token streams at 12.5 Hz → 48 kHz stereo reconstruction.
|
| 15 |
|
| 16 |
-
|
| 17 |
-
[HeartMuLa/heartlib
|
| 18 |
-
|
| 19 |
-
|
|
|
|
| 20 |
|
| 21 |
## Installation and reconstruction
|
| 22 |
|
| 23 |
Use Python 3.10, a CUDA-compatible PyTorch installation, and `ffmpeg` for MP3
|
| 24 |
-
output. Install
|
| 25 |
-
do not install both the public and development copies into one environment.
|
| 26 |
|
| 27 |
```bash
|
| 28 |
-
git clone https://github.com/HeartMuLa/heartlib
|
| 29 |
-
cd heartlib
|
| 30 |
pip install -e .
|
| 31 |
-
hf auth login
|
| 32 |
hf download HeartMuLa/HeartCodec-oss-encoder --local-dir ./ckpt/HeartCodec-oss-encoder
|
| 33 |
hf download HeartMuLa/HeartCodec-oss-20260123 --local-dir ./ckpt/HeartCodec-oss-20260123
|
| 34 |
python examples/run_music_reconstruction.py \
|
| 35 |
--encoder_path ./ckpt/HeartCodec-oss-encoder \
|
| 36 |
--decoder_path ./ckpt/HeartCodec-oss-20260123 \
|
| 37 |
-
--input_path /
|
| 38 |
--save_path ./reconstructed.mp3
|
| 39 |
```
|
| 40 |
|
| 41 |
-
|
| 42 |
320 kbps, without an intermediate PCM WAV. WAV output is available by selecting
|
| 43 |
a `.wav` output path. Output is trimmed to the input duration.
|
| 44 |
|
|
|
|
| 13 |
which supplies the shared RVQ quantizer and audio decoder. Together they support
|
| 14 |
audio → eight token streams at 12.5 Hz → 48 kHz stereo reconstruction.
|
| 15 |
|
| 16 |
+
The matching code and reconstruction examples are in
|
| 17 |
+
[HeartMuLa/heartlib](https://github.com/HeartMuLa/heartlib).
|
| 18 |
+
|
| 19 |
+
You can now encode your own audio into tokens for fine-tuning HeartMuLa.
|
| 20 |
+
We hope this contribution will be useful to the music research community.
|
| 21 |
|
| 22 |
## Installation and reconstruction
|
| 23 |
|
| 24 |
Use Python 3.10, a CUDA-compatible PyTorch installation, and `ffmpeg` for MP3
|
| 25 |
+
output. Install `heartlib` in a separate environment.
|
|
|
|
| 26 |
|
| 27 |
```bash
|
| 28 |
+
git clone https://github.com/HeartMuLa/heartlib.git
|
| 29 |
+
cd heartlib
|
| 30 |
pip install -e .
|
|
|
|
| 31 |
hf download HeartMuLa/HeartCodec-oss-encoder --local-dir ./ckpt/HeartCodec-oss-encoder
|
| 32 |
hf download HeartMuLa/HeartCodec-oss-20260123 --local-dir ./ckpt/HeartCodec-oss-20260123
|
| 33 |
python examples/run_music_reconstruction.py \
|
| 34 |
--encoder_path ./ckpt/HeartCodec-oss-encoder \
|
| 35 |
--decoder_path ./ckpt/HeartCodec-oss-20260123 \
|
| 36 |
+
--input_path ./assets/reference.mp3 \
|
| 37 |
--save_path ./reconstructed.mp3
|
| 38 |
```
|
| 39 |
|
| 40 |
+
Replace `--input_path` to use your own authorized audio. MP3 is encoded directly from float32 output at
|
| 41 |
320 kbps, without an intermediate PCM WAV. WAV output is available by selecting
|
| 42 |
a `.wav` output path. Output is trimmed to the input duration.
|
| 43 |
|