jahid
commited on
Commit
·
bc4f968
1
Parent(s):
66d626c
Initial upload of XCodec2 retrained model
Browse files
LICENSE
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
CC-BY-NC-4.0
|
README.md
ADDED
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: cc-by-nc-4.0
|
3 |
+
tags:
|
4 |
+
- audio
|
5 |
+
- codec
|
6 |
+
- speech
|
7 |
+
- xcodec2
|
8 |
+
- text-to-speech
|
9 |
+
- multilingual
|
10 |
+
language:
|
11 |
+
- en
|
12 |
+
- ja
|
13 |
+
- zh
|
14 |
+
- bn
|
15 |
+
- fr
|
16 |
+
- de
|
17 |
+
- ko
|
18 |
+
---
|
19 |
+
|
20 |
+
# 🗣️ XCodec2 Retrained (Multilingual, 100 Hours)
|
21 |
+
|
22 |
+
This model is a retrained version of [HKUSTAudio/xcodec2](https://huggingface.co/HKUSTAudio/xcodec2), trained on a 100K-hour multilingual dataset across 7 languages. It is optimized for speech representation learning, compression, and high-fidelity reconstruction — particularly useful for TTS and bandwidth-efficient speech synthesis.
|
23 |
+
|
24 |
+
---
|
25 |
+
|
26 |
+
## 📌 Overview
|
27 |
+
|
28 |
+
- **Base Model:** [HKUSTAudio/xcodec2](https://huggingface.co/HKUSTAudio/xcodec2)
|
29 |
+
- **Sampling Rate:** 16 kHz
|
30 |
+
- **Tokens:** 50 tokens/second
|
31 |
+
- **Developed By:** [Verbex.ai (Hishab Technologies Ltd.)](https://verbex.ai)
|
32 |
+
- **Primary Use Case:** High-quality speech reconstruction and intermediate TTS representations
|
33 |
+
- **Training Time:** 11 Days
|
34 |
+
- **Epoch:** 1
|
35 |
+
- **Compute:** 8xH100 80GB
|
36 |
+
|
37 |
+
---
|
38 |
+
|
39 |
+
## 🧪 Installation & Usage
|
40 |
+
|
41 |
+
This model requires `xcodec2`. We recommend using a minimal setup:
|
42 |
+
|
43 |
+
```bash
|
44 |
+
# Create environment
|
45 |
+
conda create -n xcodec2 python=3.9
|
46 |
+
conda activate xcodec2
|
47 |
+
|
48 |
+
# Install xcodec2 (choose one)
|
49 |
+
pip install xcodec2==0.1.5 # Modified, fewer dependencies (recommended for inference and LLASA fine-tuning)
|
50 |
+
# OR
|
51 |
+
pip install xcodec2==0.1.3 # Original, more stable during training
|
52 |
+
```
|
53 |
+
|
54 |
+
### Example Usage
|
55 |
+
|
56 |
+
```python
|
57 |
+
import torch
|
58 |
+
import soundfile as sf
|
59 |
+
from xcodec2.modeling_xcodec2 import XCodec2Model
|
60 |
+
|
61 |
+
model_path = "hishab/titu-xcodec2" # Replace with actual Hugging Face path
|
62 |
+
model = XCodec2Model.from_pretrained(model_path)
|
63 |
+
model.eval().cuda()
|
64 |
+
|
65 |
+
# Load and preprocess waveform
|
66 |
+
wav, sr = sf.read("test_bn.wav")
|
67 |
+
if sr != 16000:
|
68 |
+
import librosa
|
69 |
+
wav = librosa.resample(wav, orig_sr=sr, target_sr=16000)
|
70 |
+
sr = 16000
|
71 |
+
if len(wav.shape) > 1:
|
72 |
+
wav = wav.mean(axis=1)
|
73 |
+
wav_tensor = torch.from_numpy(wav).float().unsqueeze(0)
|
74 |
+
|
75 |
+
# Encode and decode
|
76 |
+
with torch.no_grad():
|
77 |
+
vq_code = model.encode_code(input_waveform=wav_tensor)
|
78 |
+
print("Code:", vq_code)
|
79 |
+
|
80 |
+
recon_wav = model.decode_code(vq_code).cpu()
|
81 |
+
|
82 |
+
# Save output
|
83 |
+
sf.write("reconstructed_bn.wav", recon_wav[0, 0].numpy(), sr)
|
84 |
+
print("Done! Check reconstructed_bn.wav")
|
85 |
+
```
|
86 |
+
|
87 |
+
---
|
88 |
+
|
89 |
+
## 🌍 Multilingual Training Dataset
|
90 |
+
|
91 |
+
| Language | Dataset(s) | Hours (K) |
|
92 |
+
|-----------|----------------------------------------|-----------|
|
93 |
+
| Japanese | EmiliaYODAS + Verbex JA TTS Dataset | 31.41 |
|
94 |
+
| English | EmiliaYODAS | 25.69 |
|
95 |
+
| Chinese | EmiliaYODAS | 12.50 |
|
96 |
+
| Bangla | Verbex Bengali TTS Dataset | 11.58 |
|
97 |
+
| French | EmiliaYODAS + MLangLibrispeech | 8.40 |
|
98 |
+
| German | EmiliaYODAS + MLangLibrispeech | 5.42 |
|
99 |
+
| Korean | EmiliaYODAS | 5.00 |
|
100 |
+
| **Total** | — | **100** |
|
101 |
+
|
102 |
+
---
|
103 |
+
|
104 |
+
## 📊 Reconstruction Evaluation
|
105 |
+
|
106 |
+
Reconstruction metrics are computed over 100 samples for English, Japanese, and Bangla using this retrained model (`XCODEC2 Ours`) alongside baselines (XCODEC, SNAC, NEMO).
|
107 |
+
|
108 |
+
**Evaluation Test Sets:**
|
109 |
+
- English: 100 Examples (Emilia Dataset @ 24 kHz)
|
110 |
+
- Japanese: 100 Examples (Emilia Dataset @ 24 kHz)
|
111 |
+
- Bangla: 100 Examples (Inhouse TTS Dataset @ 22.05 kHz)
|
112 |
+
|
113 |
+
| Model | Lang | MCD ↓ | MSE ↑ | BERTScore ↑ | BLEU ↑ | TokenDist ↑ |
|
114 |
+
|-------------------|------|--------|--------|-------------|--------|-------------|
|
115 |
+
| **XCODEC** | BN | 2.823 | 0.003 | 0.939 | 0.500 | 0.816 |
|
116 |
+
| | EN | 3.166 | 0.012 | 0.962 | 0.660 | 0.856 |
|
117 |
+
| | JA | 3.021 | 0.010 | 0.948 | 0.582 | 0.838 |
|
118 |
+
| **Overall** | | 3.003 | 0.008 | 0.949 | 0.581 | 0.837 |
|
119 |
+
| **XCODEC2 (Ours)** | BN | 2.712 | 0.003 | 0.940 | 0.508 | 0.817 |
|
120 |
+
| | EN | 3.206 | 0.014 | 0.957 | 0.644 | 0.851 |
|
121 |
+
| | JA | 3.022 | 0.012 | 0.946 | 0.573 | 0.838 |
|
122 |
+
| **Overall** | | 2.980 | 0.010 | 0.948 | 0.575 | 0.835 |
|
123 |
+
| **hubertsiuzdak/snac_24khz** | BN | 3.104 | 0.002 | 0.911 | 0.442 | 0.785 |
|
124 |
+
| | EN | 3.983 | 0.014 | 0.912 | 0.541 | 0.797 |
|
125 |
+
| | JA | 3.512 | 0.009 | 0.903 | 0.472 | 0.761 |
|
126 |
+
| **Overall** | | 3.533 | 0.008 | 0.909 | 0.485 | 0.781 |
|
127 |
+
| **nvidia/low-frame-rate-speech-codec-22khz** | BN | 2.247 | 0.000 | 0.957 | 0.589 | 0.863 |
|
128 |
+
| | EN | 2.867 | 0.007 | 0.969 | 0.707 | 0.872 |
|
129 |
+
| | JA | 2.677 | 0.003 | 0.955 | 0.614 | 0.853 |
|
130 |
+
| **Overall** | | 2.597 | 0.003 | 0.960 | 0.636 | 0.863 |
|
131 |
+
|
132 |
+
---
|
133 |
+
|
134 |
+
## ✅ Intended Use
|
135 |
+
|
136 |
+
This model is suitable for:
|
137 |
+
|
138 |
+
- Speech tokenization in TTS pipelines
|
139 |
+
- Low-bitrate speech compression
|
140 |
+
- Representation learning and fine-tuning (e.g., LLASA-style)
|
141 |
+
- Code-based speech synthesis or generation tasks
|
142 |
+
|
143 |
+
---
|
144 |
+
|
145 |
+
## 🚫 Limitations
|
146 |
+
- Licensed for **non-commercial use only**
|
147 |
+
|
148 |
+
---
|
149 |
+
|
150 |
+
## 📄 License
|
151 |
+
|
152 |
+
This model is licensed under **Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0)**.
|
153 |
+
Commercial usage is **not allowed**.
|
154 |
+
|
155 |
+
- SPDX Identifier: `CC-BY-NC-4.0`
|
156 |
+
- License Details: [https://creativecommons.org/licenses/by-nc/4.0](https://creativecommons.org/licenses/by-nc/4.0)
|
157 |
+
|
158 |
+
---
|
159 |
+
|
160 |
+
## 📬 Contact
|
161 |
+
|
162 |
+
For research collaborations, feedback, or commercial licensing inquiries, please reach out to:
|
163 |
+
|
164 |
+
**📧 Email:** [[email protected]]
|
165 |
+
**🌐 Website:** [https://verbex.ai](https://verbex.ai)
|
166 |
+
---
|