README.md

---
language:
  - vi
license: cc-by-nc-sa-4.0
multilinguality: monolingual
size_categories:
  - n<1K
source_datasets:
  - linhtran92/viet_bud500
task_categories:
  - automatic-speech-recognition
tags:
  - vietnamese
  - benchmark
  - quantization
  - asr
  - wer
  - cer
  - calibration
  - evaluation
  - audio
  - speech
pretty_name: ViASR-Bench
dataset_info:
  features:
    - name: id
      dtype: string
    - name: transcript
      dtype: string
    - name: duration_s
      dtype: float32
    - name: sampling_rate
      dtype: int32
    - name: source_split
      dtype: string
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
  splits:
    - name: calibration
      num_examples: 256
    - name: evaluation
      num_examples: 600
  download_size: 73400320
  dataset_size: 73400320
---

# ViASR-Bench 🇻🇳🎙️

**Vietnamese benchmark dataset for ASR model quantization — providing both a calibration set and an evaluation set.**

ViASR-Bench is a curated sample from [BUD500](https://huggingface.co/datasets/linhtran92/viet_bud500),
the largest public Vietnamese podcast speech dataset.
It is designed specifically for the **two-phase lifecycle of ASR quantization**:

| Phase | Split | Purpose |
|-------|-------|---------|
| **Calibration** | `calibration` (256 utterances) | Feed to AWQ / GPTQ / SmoothQuant to compute activation statistics and compensation factors |
| **Evaluation**  | `evaluation` (600 utterances)  | Measure WER / CER degradation after quantization |

> **Important:** This dataset is a **subsample of BUD500**, not a standalone corpus.
> Audio content is sourced from BUD500 (CC-BY-NC-SA 4.0) and is intended for
> **non-commercial research use only**.

---

## Dataset Summary

| Split | Utterances | Duration | Size (~) | Source in BUD500 |
|-------|-----------|----------|----------|------------------|
| `calibration` | 256 | ~10.9 min | ~21 MB | `train` split |
| `evaluation`  | 600 | ~25.6 min | ~49 MB | `test` split  |
| **Total**     | **856** | **~36.5 min** | **~70 MB** | |

**Audio format:** 16 kHz · Mono · PCM-16 (WAV)

**Utterance duration statistics:**

| Split | Mean (s) | Median (s) | P10 (s) | P90 (s) | Max (s) |
|-------|---------|-----------|---------|---------|---------|
| calibration | 2.56 | 2.12 | 1.10 | 4.48 | 19.2 |
| evaluation  | 2.56 | 2.11 | 1.09 | 4.52 | 19.8 |

---

## Motivation

### Why a Vietnamese-specific ASR quantization benchmark?

Existing quantization tools (AWQ, GPTQ, SmoothQuant) require a **calibration dataset**
whose acoustic distribution matches the target language. Using English speech data
to calibrate a Vietnamese ASR model leads to suboptimal quantization because:

1. **6-tone tonal system.** Vietnamese has six tones (ngang, huyền, sắc, hỏi, ngã, nặng)
   that create unique patterns in mel-spectrograms and encoder feature spaces,
   distinct from non-tonal languages.

2. **Regional dialect variation.** Three major dialects (Northern, Central, Southern)
   differ significantly in formant distributions and prosody, affecting attention
   and convolution layers in ASR encoders.

3. **Monosyllabic structure.** Vietnamese is predominantly monosyllabic, producing
   different sequence-length and attention-pattern distributions than English.

Calibrating on English audio leaves these activation patterns unrepresented,
degrading quantization quality specifically for Vietnamese phonemes.

### Why CER over WER for Vietnamese?

Vietnamese words average **1.7–2.2 characters**, making tone confusion errors
(e.g., *khoẻ* → *khoề*: a single diacritic change) inflate WER disproportionately:

```
Reference:  "bạn có khoẻ không"   (4 words, 15 chars)
Hypothesis: "bạn có khoề không"

WER = 1/4 = 25.0%   ← one wrong word
CER = 1/15 = 6.7%   ← one wrong character
```

CER is a more sensitive and fair metric for detecting tone-level degradation
caused by quantization on Vietnamese. **Both WER and CER are reported.**

---

## Source Data: BUD500

BUD500 is the largest public Vietnamese speech dataset (~500 hours of podcast audio
with manually verified transcripts, covering diverse topics and all three major dialects).

| Property | Value |
|----------|-------|
| Total duration | ~500 hours |
| Utterances | ~150,000 |
| Sampling rate | 16,000 Hz |
| Content type | Podcast (multi-topic) |
| Dialect coverage | Northern, Central, Southern |
| HuggingFace ID | `linhtran92/viet_bud500` |
| License | CC-BY-NC-SA 4.0 |

**Why BUD500 over alternatives?**

| Dataset | Hours | Quality | Diversity | Public |
|---------|-------|---------|-----------|--------|
| **BUD500** | ~500h | High | High | ✅ |
| VIVOS | 15h | High | Low (read speech) | ✅ |
| FLEURS vi | ~10h | Medium | Medium | ✅ |
| CommonVoice vi | ~30h | Medium | Medium | ✅ |
| VLSP | 100h+ | High | High | ❌ |

BUD500's size ensures that random sampling still yields diverse coverage.
Its podcast format is closer to real-world deployment conditions than
studio read-speech datasets like VIVOS.

---

## Sampling Methodology

### Non-overlapping splits

Calibration and evaluation sets are **guaranteed non-overlapping by transcript**:

```
eval_transcripts = {s["transcript"] for s in eval_samples}

# Calibration: skip any utterance whose transcript appears in eval
if transcript in eval_transcripts:
    continue
```

This prevents data leakage between calibration and evaluation.

### Quality filter

Each utterance passes a 2-condition filter:

```
keep(u) = True  iff:
  1.0s ≤ duration(u) ≤ 20.0s     # remove noise clips and unusually long segments
  AND  len(transcript(u)) > 0     # non-empty transcript
```

### Sampling procedure

```
1. Stream BUD500 (streaming=True) — no full download required
2. Collect buffer = 5× target size from filtered utterances
3. random.sample(buffer, n_target, seed=42)  — fixed seed for reproducibility
```

**Calibration** is drawn from BUD500 `train` split (seed=43).
**Evaluation** is drawn from BUD500 `test` split (seed=42).

### Transcript normalization

All transcripts are stored in **Unicode NFC** normalized form.
Before computing WER/CER, apply the same normalization to model hypotheses:

```python
import unicodedata, re

def normalise_vi(text: str) -> str:
    text = unicodedata.normalize("NFC", text)   # critical for Vietnamese diacritics
    text = text.lower()
    text = re.sub(r"[^\w\s]", " ", text, flags=re.UNICODE)
    text = re.sub(r"\s+", " ", text).strip()
    return text
```

NFC normalization is critical because Vietnamese characters can be encoded two ways:
- Decomposed: `e` + combining hook above + combining dot below
- Precomposed: single codepoint `ệ`

Without normalization, visually identical strings may have non-zero edit distance.

---

## Usage

### Installation

```bash
pip install datasets soundfile jiwer transformers torch tqdm numpy
```

### Load the dataset

```python
from datasets import load_dataset

# Evaluation split
eval_ds = load_dataset("your-org/vi-asr-bench", split="evaluation")

# Calibration split
calib_ds = load_dataset("your-org/vi-asr-bench", split="calibration")

# Each sample:
# {
#   "id":            "sample_0042",
#   "audio":         {"array": np.ndarray, "sampling_rate": 16000},
#   "transcript":    "xin chào thế giới",
#   "duration_s":    3.21,
#   "source_split":  "test"
# }
```

### AWQ / GPTQ Calibration

```python
import torch
from datasets import load_dataset

calib_ds = load_dataset("your-org/vi-asr-bench", split="calibration")

# Extract audio arrays for the encoder
audio_inputs = [
    torch.tensor(sample["audio"]["array"]).unsqueeze(0)   # (1, T)
    for sample in calib_ds
]

# Feed into your quantization calibration loop:
# for audio in audio_inputs:
#     model.encoder(audio)   # collect activation statistics
```

### WER / CER Evaluation

```python
import unicodedata, re
from datasets import load_dataset
from jiwer import wer, cer
from transformers import pipeline

def normalise_vi(text):
    text = unicodedata.normalize("NFC", text)
    text = text.lower()
    text = re.sub(r"[^\w\s]", " ", text, flags=re.UNICODE)
    return re.sub(r"\s+", " ", text).strip()

eval_ds  = load_dataset("your-org/vi-asr-bench", split="evaluation")
asr_pipe = pipeline("automatic-speech-recognition",
                    model="your-quantized-model",
                    generate_kwargs={"language": "vi", "task": "transcribe"})

references  = [normalise_vi(s["transcript"]) for s in eval_ds]
hypotheses  = [normalise_vi(r["text"]) for r in asr_pipe(
                   [s["audio"] for s in eval_ds], batch_size=8)]

print(f"WER: {wer(references, hypotheses)*100:.2f}%")
print(f"CER: {cer(references, hypotheses)*100:.2f}%")
```

### Compare quantized vs. baseline

```bash
python vi_asr_eval.py \
    --model   openai/whisper-large-v3 \
    --compare your-org/whisper-large-v3-int4 \
    --output  results.json
```

Expected output format:

```
================================================================
Metric        openai/whisper-large-v3   your-org/...-int4    Delta
----------------------------------------------------------------
WER                          8.24%              8.91%       +0.67%
CER                          3.11%              3.38%       +0.27%
```

---

## Reproducibility

Full metadata is stored in `metadata.json` for each split:

```json
{
  "split": "evaluation",
  "n_samples": 600,
  "seed": 42,
  "source": "linhtran92/viet_bud500",
  "source_split": "test",
  "sampling_rate": 16000,
  "total_duration_s": 1536.0,
  "total_duration_min": 25.6,
  "total_size_mb": 49.1,
  "filter": {"min_dur_s": 1.0, "max_dur_s": 20.0},
  "normalization": "unicode_nfc"
}
```

---

## Relation to ViWiki-Bench

ViASR-Bench is the **speech counterpart** of ViWiki-Bench (Vietnamese LLM benchmark).
Together they form a complete Vietnamese quantization benchmark suite:

| | ViWiki-Bench (LLM) | ViASR-Bench (ASR) |
|---|---|---|
| Input modality | Text (token IDs) | Audio (float32 arrays) |
| Calibration | 128 sequences from Wikipedia vi | 256 utterances from BUD500 train |
| Evaluation | ~280k tokens from Wikipedia vi | 600 utterances from BUD500 test |
| Metric | Perplexity (PPL) | WER / CER |
| Source dataset | `wikimedia/wikipedia` 20231101.vi | `linhtran92/viet_bud500` |

---

## Limitations

- **Single source:** Only BUD500 (podcast). Read speech (VIVOS), spontaneous conversation,
  and children's speech are not represented.
- **No dialect labels:** BUD500 does not tag individual utterances with Northern/Central/Southern
  dialect, so stratified sampling by dialect is not possible.
- **Small evaluation set:** 600 utterances (~26 min) is sufficient for relative comparison
  between quantization methods, but not for high-precision absolute WER estimation
  (vs. LibriSpeech test-clean at 5.4 hours).
- **Non-commercial license:** Inherited from BUD500 (CC-BY-NC-SA 4.0).
  Not suitable for commercial product deployment.

---

## Citation

If you use ViASR-Bench in your research, please cite both this dataset and BUD500:

```bibtex
@techreport{viasr_bench2024,
  title     = {ViASR-Bench: A Vietnamese Benchmark Dataset for
               ASR Model Quantization Calibration and Evaluation},
  author    = {AnhND},
  year      = {2026},
  note      = {Technical Report v1.0. Sampled from linhtran92/viet\_bud500},
  url       = {https://huggingface.co/datasets/your-org/vi-asr-bench}
}

@dataset{bud500,
  title     = {BUD500: A Large-Scale Vietnamese Podcast Speech Dataset},
  author    = {Linh Tran},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/linhtran92/viet_bud500}
}
```

---

## License

This dataset inherits the license of its source:
**CC-BY-NC-SA 4.0** (Non-Commercial).

The sampling and evaluation code is released under **MIT License**.