Update README from MODEL_CARD.md
Browse files
README.md
CHANGED
|
@@ -1,142 +1,267 @@
|
|
| 1 |
---
|
| 2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
tags:
|
| 4 |
-
-
|
| 5 |
-
|
| 6 |
-
-
|
| 7 |
-
|
|
|
|
|
|
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
#
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
datasets:
|
| 6 |
+
- speechbrain/LoquaciousSet
|
| 7 |
+
base_model:
|
| 8 |
+
- zai-org/GLM-ASR-Nano-2512
|
| 9 |
+
- Qwen/Qwen3-0.6B
|
| 10 |
+
pipeline_tag: automatic-speech-recognition
|
| 11 |
tags:
|
| 12 |
+
- asr
|
| 13 |
+
- speech-recognition
|
| 14 |
+
- audio
|
| 15 |
+
- qwen
|
| 16 |
+
- glm-asr
|
| 17 |
+
library_name: transformers
|
| 18 |
---
|
| 19 |
|
| 20 |
+
# Tiny Audio
|
| 21 |
+
|
| 22 |
+
A speech recognition model trained in 24 hours on a single GPU for ~$12. Built with [Tiny Audio](https://github.com/alexkroman/tiny-audio)—a minimal, hackable ASR framework.
|
| 23 |
+
|
| 24 |
+
## Quick Start
|
| 25 |
+
|
| 26 |
+
```python
|
| 27 |
+
from transformers import pipeline
|
| 28 |
+
|
| 29 |
+
pipe = pipeline("automatic-speech-recognition", model="mazesmazes/tiny-audio", trust_remote_code=True)
|
| 30 |
+
result = pipe("audio.wav")
|
| 31 |
+
print(result["text"])
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
## Usage Examples
|
| 35 |
+
|
| 36 |
+
### Basic Transcription
|
| 37 |
+
|
| 38 |
+
```python
|
| 39 |
+
from transformers import pipeline
|
| 40 |
+
|
| 41 |
+
pipe = pipeline("automatic-speech-recognition", model="mazesmazes/tiny-audio", trust_remote_code=True)
|
| 42 |
+
|
| 43 |
+
# From file
|
| 44 |
+
result = pipe("audio.wav")
|
| 45 |
+
print(result["text"])
|
| 46 |
+
|
| 47 |
+
# From URL
|
| 48 |
+
result = pipe("https://example.com/audio.mp3")
|
| 49 |
+
|
| 50 |
+
# From numpy array (must be 16kHz)
|
| 51 |
+
import numpy as np
|
| 52 |
+
audio = np.random.randn(16000).astype(np.float32) # 1 second
|
| 53 |
+
result = pipe(audio)
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
### Batch Processing
|
| 57 |
+
|
| 58 |
+
```python
|
| 59 |
+
# Process multiple files
|
| 60 |
+
files = ["audio1.wav", "audio2.wav", "audio3.wav"]
|
| 61 |
+
results = pipe(files, batch_size=4)
|
| 62 |
+
for r in results:
|
| 63 |
+
print(r["text"])
|
| 64 |
+
```
|
| 65 |
+
|
| 66 |
+
### Word-Level Timestamps
|
| 67 |
+
|
| 68 |
+
```python
|
| 69 |
+
result = pipe("audio.wav", return_timestamps="word")
|
| 70 |
+
# Returns:
|
| 71 |
+
# {
|
| 72 |
+
# "text": "hello world",
|
| 73 |
+
# "chunks": [
|
| 74 |
+
# {"text": "hello", "timestamp": (0.0, 0.5)},
|
| 75 |
+
# {"text": "world", "timestamp": (0.6, 1.0)}
|
| 76 |
+
# ]
|
| 77 |
+
# }
|
| 78 |
+
```
|
| 79 |
+
|
| 80 |
+
### Streaming Inference
|
| 81 |
+
|
| 82 |
+
```python
|
| 83 |
+
from tiny_audio import ASRModel, ASRProcessor
|
| 84 |
+
import torch
|
| 85 |
+
|
| 86 |
+
model = ASRModel.from_pretrained("mazesmazes/tiny-audio")
|
| 87 |
+
processor = ASRProcessor.from_pretrained("mazesmazes/tiny-audio")
|
| 88 |
+
|
| 89 |
+
# Load and process audio
|
| 90 |
+
import librosa
|
| 91 |
+
audio, sr = librosa.load("audio.wav", sr=16000)
|
| 92 |
+
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
|
| 93 |
+
|
| 94 |
+
# Stream tokens
|
| 95 |
+
for token in model.generate_streaming(inputs["input_features"]):
|
| 96 |
+
print(token, end="", flush=True)
|
| 97 |
+
```
|
| 98 |
+
|
| 99 |
+
### Using with torch directly
|
| 100 |
+
|
| 101 |
+
```python
|
| 102 |
+
from tiny_audio import ASRModel, ASRProcessor
|
| 103 |
+
import torch
|
| 104 |
+
import librosa
|
| 105 |
+
|
| 106 |
+
# Load model and processor
|
| 107 |
+
model = ASRModel.from_pretrained("mazesmazes/tiny-audio")
|
| 108 |
+
processor = ASRProcessor.from_pretrained("mazesmazes/tiny-audio")
|
| 109 |
+
|
| 110 |
+
# Load audio (16kHz)
|
| 111 |
+
audio, sr = librosa.load("audio.wav", sr=16000)
|
| 112 |
+
|
| 113 |
+
# Process
|
| 114 |
+
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
|
| 115 |
+
|
| 116 |
+
# Generate
|
| 117 |
+
with torch.no_grad():
|
| 118 |
+
output = model.generate(
|
| 119 |
+
input_features=inputs["input_features"],
|
| 120 |
+
attention_mask=inputs["attention_mask"],
|
| 121 |
+
max_new_tokens=256
|
| 122 |
+
)
|
| 123 |
+
|
| 124 |
+
# Decode
|
| 125 |
+
text = processor.batch_decode(output, skip_special_tokens=True)[0]
|
| 126 |
+
print(text)
|
| 127 |
+
```
|
| 128 |
+
|
| 129 |
+
### GPU Inference
|
| 130 |
+
|
| 131 |
+
```python
|
| 132 |
+
import torch
|
| 133 |
+
|
| 134 |
+
pipe = pipeline(
|
| 135 |
+
"automatic-speech-recognition",
|
| 136 |
+
model="mazesmazes/tiny-audio",
|
| 137 |
+
trust_remote_code=True,
|
| 138 |
+
device="cuda" # or device=0
|
| 139 |
+
)
|
| 140 |
+
```
|
| 141 |
+
|
| 142 |
+
### Half Precision
|
| 143 |
+
|
| 144 |
+
```python
|
| 145 |
+
pipe = pipeline(
|
| 146 |
+
"automatic-speech-recognition",
|
| 147 |
+
model="mazesmazes/tiny-audio",
|
| 148 |
+
trust_remote_code=True,
|
| 149 |
+
torch_dtype=torch.float16,
|
| 150 |
+
device="cuda"
|
| 151 |
+
)
|
| 152 |
+
```
|
| 153 |
+
|
| 154 |
+
## Architecture
|
| 155 |
+
|
| 156 |
+
```
|
| 157 |
+
Audio (16kHz) → GLM-ASR Encoder (frozen) → MLP Projector (trained) → Qwen3 (frozen) → Text
|
| 158 |
+
```
|
| 159 |
+
|
| 160 |
+
Only the projector is trained (~12M params). The encoder and decoder remain frozen, leveraging their pretrained knowledge.
|
| 161 |
+
|
| 162 |
+
| Component | Model | Parameters | Status |
|
| 163 |
+
|-----------|-------|------------|--------|
|
| 164 |
+
| Audio Encoder | GLM-ASR-Nano-2512 | ~600M | Frozen |
|
| 165 |
+
| Projector | 2-layer MLP | ~12M | Trained |
|
| 166 |
+
| Language Model | Qwen3-0.6B | ~600M | Frozen |
|
| 167 |
+
|
| 168 |
+
### How It Works
|
| 169 |
+
|
| 170 |
+
1. **Audio Encoder**: GLM-ASR converts 16kHz audio into frame-level embeddings (768-dim)
|
| 171 |
+
2. **Projector**: A 2-layer MLP with frame stacking bridges the audio and text embedding spaces
|
| 172 |
+
3. **Language Model**: Qwen3 generates text autoregressively, conditioned on the projected audio
|
| 173 |
+
|
| 174 |
+
The projector reduces sequence length via frame stacking: `output_len = (input_len - 5) // 5 + 1`
|
| 175 |
+
|
| 176 |
+
## Model Specifications
|
| 177 |
+
|
| 178 |
+
| Specification | Value |
|
| 179 |
+
|---------------|-------|
|
| 180 |
+
| Input | Audio (16kHz mono) |
|
| 181 |
+
| Output | Text transcription |
|
| 182 |
+
| Max Audio Length | ~30 seconds (limited by encoder) |
|
| 183 |
+
| Vocabulary | Qwen3 tokenizer |
|
| 184 |
+
| Languages | English only |
|
| 185 |
+
| Generation | Greedy decoding (num_beams=1, do_sample=False) |
|
| 186 |
+
|
| 187 |
+
## Training Details
|
| 188 |
+
|
| 189 |
+
| | |
|
| 190 |
+
|---|---|
|
| 191 |
+
| **Dataset** | LoquaciousSet (25,000 hours) |
|
| 192 |
+
| **Hardware** | Single NVIDIA A40 |
|
| 193 |
+
| **Time** | ~24 hours |
|
| 194 |
+
| **Cost** | ~$12 |
|
| 195 |
+
| **Optimizer** | AdamW |
|
| 196 |
+
| **Learning Rate** | 1e-4 |
|
| 197 |
+
| **Batch Size** | 4 |
|
| 198 |
+
| **Steps** | 50,000 |
|
| 199 |
+
|
| 200 |
+
## Limitations
|
| 201 |
+
|
| 202 |
+
- **English only**: Not trained on other languages
|
| 203 |
+
- **Sample rate**: Expects 16kHz audio (other rates resampled automatically)
|
| 204 |
+
- **Audio length**: Best for clips under 30 seconds
|
| 205 |
+
- **Accuracy**: May degrade on:
|
| 206 |
+
- Heavily accented speech
|
| 207 |
+
- Noisy or low-quality audio
|
| 208 |
+
- Domain-specific terminology
|
| 209 |
+
- Overlapping speakers
|
| 210 |
+
- **No punctuation**: Output is lowercase without punctuation by default
|
| 211 |
+
|
| 212 |
+
## Requirements
|
| 213 |
+
|
| 214 |
+
```
|
| 215 |
+
transformers>=4.40.0
|
| 216 |
+
torch>=2.0.0
|
| 217 |
+
torchaudio>=2.0.0
|
| 218 |
+
```
|
| 219 |
+
|
| 220 |
+
Optional for streaming:
|
| 221 |
+
```
|
| 222 |
+
librosa
|
| 223 |
+
soundfile
|
| 224 |
+
```
|
| 225 |
+
|
| 226 |
+
## Files
|
| 227 |
+
|
| 228 |
+
| File | Description |
|
| 229 |
+
|------|-------------|
|
| 230 |
+
| `config.json` | Model configuration |
|
| 231 |
+
| `model.safetensors` | Projector weights (~48MB) |
|
| 232 |
+
| `preprocessor_config.json` | Audio preprocessing config |
|
| 233 |
+
| `tokenizer.json` | Tokenizer |
|
| 234 |
+
| `tokenizer_config.json` | Tokenizer config |
|
| 235 |
+
| `special_tokens_map.json` | Special tokens |
|
| 236 |
+
|
| 237 |
+
Note: Only the projector weights are stored. The encoder (GLM-ASR) and decoder (Qwen3) are loaded from their respective HuggingFace repos.
|
| 238 |
+
|
| 239 |
+
## Citation
|
| 240 |
+
|
| 241 |
+
If you use this model, please cite:
|
| 242 |
+
|
| 243 |
+
```bibtex
|
| 244 |
+
@misc{tinyaudio2024,
|
| 245 |
+
author = {Alex Kroman},
|
| 246 |
+
title = {Tiny Audio: Minimal ASR Training},
|
| 247 |
+
year = {2024},
|
| 248 |
+
publisher = {GitHub},
|
| 249 |
+
url = {https://github.com/alexkroman/tiny-audio}
|
| 250 |
+
}
|
| 251 |
+
```
|
| 252 |
+
|
| 253 |
+
## Links
|
| 254 |
+
|
| 255 |
+
- [GitHub Repository](https://github.com/alexkroman/tiny-audio) - Train your own model
|
| 256 |
+
- [Free 3.5-hour Course](https://github.com/alexkroman/tiny-audio/blob/main/docs/course/0-course-overview.md) - Learn ASR from scratch
|
| 257 |
+
- [Live Demo](https://huggingface.co/spaces/mazesmazes/tiny-audio) - Try it in your browser
|
| 258 |
+
|
| 259 |
+
## Acknowledgments
|
| 260 |
+
|
| 261 |
+
- [GLM-ASR](https://huggingface.co/zai-org/GLM-ASR-Nano-2512) for the audio encoder
|
| 262 |
+
- [Qwen3](https://huggingface.co/Qwen/Qwen3-0.6B) for the language model
|
| 263 |
+
- [LoquaciousSet](https://huggingface.co/datasets/speechbrain/LoquaciousSet) for training data
|
| 264 |
+
|
| 265 |
+
## License
|
| 266 |
+
|
| 267 |
+
MIT
|