Update README.md
Browse files
README.md
CHANGED
|
@@ -1,29 +1,29 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-nd-4.0
|
| 3 |
---
|
| 4 |
-
# Vietnamese Speech-to-Text (ASR) — ZipFormer-30M-RNNT-6000h
|
| 5 |
|
| 6 |
-
## 🔍
|
| 7 |
-
|
| 8 |
-
|
| 9 |
|
| 10 |
---
|
| 11 |
|
| 12 |
-
## ⚙️
|
| 13 |
-
- **
|
| 14 |
-
- **
|
| 15 |
-
- **
|
| 16 |
-
- **Loss Function
|
| 17 |
- **Framework:** PyTorch + k2
|
| 18 |
-
- **
|
| 19 |
|
| 20 |
---
|
| 21 |
|
| 22 |
-
## 🧠
|
| 23 |
-
|
| 24 |
|
| 25 |
-
|
|
| 26 |
-
|
| 27 |
| VLSP2020 | VLSP2021 | VLSP2022 |
|
| 28 |
| VLSP2023 | FPT | VIET_BUD500 |
|
| 29 |
| VietSpeech | FLEURS | VietMed_Labeled |
|
|
@@ -31,30 +31,39 @@ Mô hình được huấn luyện với **~6000 giờ dữ liệu tiếng Việt
|
|
| 31 |
|
| 32 |
---
|
| 33 |
|
| 34 |
-
## 🧪
|
| 35 |
|
| 36 |
-
| **Dataset** | **ZipFormer-30M-6000h** | **ChunkFormer-110M-3000h** | **PhoWhisper-Large-1.5B-800h** | **VietASR-ZipFormer-68M-
|
| 37 |
-
|
| 38 |
-
| **VLSP2023-PublicTest** | **10.40** | 16.15 | 16.83 | 14.70 |
|
| 39 |
-
| **VLSP2023-PublicTest** | **11.10** | 17.12 | 17.10 | 15.07 |
|
| 40 |
-
| **VLSP2025-PublicTest** | **7.97** | 15.55 | 16.14 | 13.55 |
|
| 41 |
-
| **VLSP2025-PublicTest** | **8.10** | 16.07 | 16.31 | 13.97 |
|
| 42 |
| **GigaSpeech2-Test** | 7.56 | 10.35 | 10.00 | **6.88** |
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
---
|
| 45 |
|
| 46 |
-
|
| 47 |
-
Đặc biệt, kiến trúc này đã **giành giải Nhất** cuộc thi **Vietnamese Language Speech Processing (VLSP)** năm **2025**.
|
| 48 |
-
Chi tiết về **dữ liệu được sử dụng**, **phương pháp huấn luyện**, **tối ưu hóa mô hình**, và **kết quả đánh giá chi tiết** được trình bày trong bài báo:
|
| 49 |
|
| 50 |
-
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
---
|
| 53 |
|
| 54 |
-
##
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
| CPU (Basic Hugging Face) | 12 giây | **0.4 giây** |
|
| 58 |
-
| GPU (RTX 3090) | 12 giây | **<0.1 giây** |
|
| 59 |
|
| 60 |
---
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-nd-4.0
|
| 3 |
---
|
| 4 |
+
# 🇻🇳 Vietnamese Speech-to-Text (ASR) — ZipFormer-30M-RNNT-6000h
|
| 5 |
|
| 6 |
+
## 🔍 Overview
|
| 7 |
+
The **Vietnamese Speech-to-Text (ASR)** model is built on the **ZipFormer architecture** — an improved variant of the Conformer — featuring only **30 million parameters** yet achieving **exceptional performance** in both speed and accuracy.
|
| 8 |
+
On CPU, the model can transcribe a **12-second audio clip in just 0.4 seconds**, significantly faster than most traditional ASR systems without requiring a GPU.
|
| 9 |
|
| 10 |
---
|
| 11 |
|
| 12 |
+
## ⚙️ Model Architecture
|
| 13 |
+
- **Architecture:** ZipFormer
|
| 14 |
+
- **Parameters:** ~30M
|
| 15 |
+
- **Language:** Vietnamese
|
| 16 |
+
- **Loss Function:** RNN-Transducer (RNNT Loss)
|
| 17 |
- **Framework:** PyTorch + k2
|
| 18 |
+
- **Optimized for:** High-speed CPU inference
|
| 19 |
|
| 20 |
---
|
| 21 |
|
| 22 |
+
## 🧠 Training Data
|
| 23 |
+
The model was trained on approximately **6000 hours of high-quality Vietnamese speech** collected from various public datasets:
|
| 24 |
|
| 25 |
+
| Dataset | | |
|
| 26 |
+
|----------|----------|----------|
|
| 27 |
| VLSP2020 | VLSP2021 | VLSP2022 |
|
| 28 |
| VLSP2023 | FPT | VIET_BUD500 |
|
| 29 |
| VietSpeech | FLEURS | VietMed_Labeled |
|
|
|
|
| 31 |
|
| 32 |
---
|
| 33 |
|
| 34 |
+
## 🧪 Evaluation Results
|
| 35 |
|
| 36 |
+
| **Dataset** | **ZipFormer-30M-6000h** | **ChunkFormer-110M-3000h** | **PhoWhisper-Large-1.5B-800h** | **VietASR-ZipFormer-68M-70k h** |
|
| 37 |
+
|--------------|--------------------------|-----------------------------|--------------------------------|---------------------------------|
|
| 38 |
+
| **VLSP2023-PublicTest (Set 1)** | **10.40** | 16.15 | 16.83 | 14.70 |
|
| 39 |
+
| **VLSP2023-PublicTest (Set 2)** | **11.10** | 17.12 | 17.10 | 15.07 |
|
| 40 |
+
| **VLSP2025-PublicTest (Set 1)** | **7.97** | 15.55 | 16.14 | 13.55 |
|
| 41 |
+
| **VLSP2025-PublicTest (Set 2)** | **8.10** | 16.07 | 16.31 | 13.97 |
|
| 42 |
| **GigaSpeech2-Test** | 7.56 | 10.35 | 10.00 | **6.88** |
|
| 43 |
|
| 44 |
+
> Lower is better (WER %)
|
| 45 |
+
|
| 46 |
+
---
|
| 47 |
+
|
| 48 |
+
## 🏆 Achievements
|
| 49 |
+
This model architecture **won First Place** in the **Vietnamese Language Speech Processing (VLSP)** competition **2025**.
|
| 50 |
+
Comprehensive details about **training data**, **optimization strategies**, **architecture improvements**, and **evaluation methodologies** are available in the paper below:
|
| 51 |
+
|
| 52 |
+
👉 [Read the full paper on Overleaf](https://www.overleaf.com/read/wjntrgchhbgv#48aa25)
|
| 53 |
+
|
| 54 |
---
|
| 55 |
|
| 56 |
+
## ⚡ Inference Speed
|
|
|
|
|
|
|
| 57 |
|
| 58 |
+
| **Device** | **Audio Length** | **Inference Time** |
|
| 59 |
+
|-------------|------------------|--------------------|
|
| 60 |
+
| CPU (Hugging Face Basic) | 12 seconds | **0.4 s** |
|
| 61 |
+
| GPU (RTX 3090) | 12 seconds | **< 0.1 s** |
|
| 62 |
|
| 63 |
---
|
| 64 |
|
| 65 |
+
## 💬 Summary
|
| 66 |
+
The **ZipFormer-30M-RNNT-6000h** model demonstrates that a lightweight architecture can still achieve state-of-the-art accuracy for Vietnamese ASR.
|
| 67 |
+
It is designed for **fast deployment on CPU-based systems**, making it ideal for **real-time speech recognition**, **callbots**, and **embedded speech interfaces**.
|
|
|
|
|
|
|
| 68 |
|
| 69 |
---
|