niloycste68/Bangali_local_dialect_ASR_HF_Dataset

BanglaMix — Dialectal Bengali–English Code-Switching ASR

The first large-scale dataset and benchmark for ASR on dialectal Bangladeshi speech mixed with English.

Dataset	41,499 clips · 131.23 hours (unaugmented) · 15 dialects
Hugging Face	niloycste68/Bangali_local_dialect_ASR_HF_Dataset
Task	Automatic Speech Recognition (ASR)
Languages	Bengali (15 regional dialects) + English code-switching
Target venues	INTERSPEECH 2026 / ACL 2026

---

Research Problem

Existing Bengali ASR resources cover either dialect or code-switching (CS) — never both:

Dataset	Dialect?	Code-Switching?	Hours
OpenSLR-37	No	No	100
MUCS 2021	No	Yes (standard Bengali only)	16
Ben-10 (AACL 2025)	Yes (10 dialects)	No	78
BanglaMix (ours)	Yes (15 dialects)	Yes	131

Young Bangladeshi speakers naturally mix English into regional dialect speech ("চাটগাঁইয়া ভাষায় phone unboxing করতাছি, honestly speaking camera performance অনেক ভালো"). No model or dataset covers this intersection.

---

Project Structure

ASR work bengali/
│
├── dataset_pipeline/          ← Phase 1: Data collection & preparation
│   ├── 01_download_audio.py       Download from YouTube (dialect-name queries)
│   ├── 01b_download_extra_urls.py Download from Facebook / other sites
│   ├── convert_webm.py            Convert .webm/.m4a → .wav (16kHz mono)
│   ├── 02_segment_audio.py        VAD segmentation → 5–15 sec clips
│   ├── 03_auto_transcribe.py      Whisper large-v3 transcription + CS detection
│   ├── 04b_auto_annotate.py       Auto-accept high-confidence transcripts
│   ├── 05_build_dataset.py        Train/dev/test split (source-isolated, dialect-aware)
│   ├── code_mixing_utils.py       CS detection utilities (word-level script tags)
│   ├── pipeline_utils.py          Shared utilities (command checking, imports)
│   └── download_query_profiles.py YouTube query profiles per dialect
│
├── fine_tuning/               ← Phase 2: Model training
│   ├── config.py                  Central config (all paths, hyperparams, model keys)
│   ├── 01_prepare_hf_dataset.py   Convert CSVs → HuggingFace Dataset format
│   ├── 02_finetune_whisper.py     Fine-tune Whisper (3 subsets: all/bn_only/cs_only)
│   └── 03_finetune_wav2vec2.py    Fine-tune wav2vec2 / WavLM (CTC models)
│
├── evaluation/                ← Phase 3: Evaluation & paper results
│   ├── 01_run_baselines.py        Zero-shot: Whisper, Tugstugi, MMS, wav2vec2
│   ├── 02_run_finetuned.py        Run all fine-tuned models on test set
│   ├── 03_compute_metrics.py      WER, CER, MIX-ER, WER-BN, WER-EN, CS-F1
│   ├── 04_ablation_summary.py     Ablation table → .csv + .tex (paste into paper)
│   ├── 05_plot_results.py         5 publication-ready plots (PDF + PNG)
│   ├── 06_statistical_analysis.py Bootstrap CI, p-values, Cohen's d
│   ├── 07_cross_dialect_eval.py   Per-dialect WER + leave-one-out experiment
│   ├── 08_error_analysis.py       Sub/Del/Ins breakdown, EN word survival
│   ├── 09_lm_shallow_fusion.py    5-gram KenLM + beam search for CTC models
│   └── 10_whisper_size_comparison.py  tiny→large-v3 accuracy vs efficiency
│
├── asr_paper_draft/           ← Phase 4: LaTeX paper
│   ├── main.tex                   Master file (compile this)
│   ├── references.bib             All citations
│   ├── Makefile                   Run: make
│   └── sections/
│       ├── abstract.tex
│       ├── 01_introduction.tex
│       ├── 02_related_work.tex
│       ├── 03_dataset.tex
│       ├── 04_methodology.tex
│       ├── 05_experiments.tex
│       ├── 06_results.tex
│       ├── 07_analysis.tex
│       └── 08_conclusion.tex
│
└── requirements.txt           ← Install everything from here

---

Installation

# 1. Create conda environment
conda create -n ASR python=3.12
conda activate ASR

# 2. Install PyTorch with CUDA (recommended — replace cu121 with your CUDA version)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121

# 3. Install all other requirements
pip install -r requirements.txt

System dependencies (install once, not via pip):

• ffmpeg — required by pydub and Whisper

  # Windows:
  winget install ffmpeg
  # or: choco install ffmpeg
  

• KenLM — required only for LM shallow fusion (evaluation Step 11)

  pip install https://github.com/kpu/kenlm/archive/master.zip
  pip install pyctcdecode
  

• Ollama — required only for LLM-as-a-Judge ensemble (evaluation Step 11 / 11_llm_judge.py)

  # 1. Install Ollama: https://ollama.com/download
  # 2. Pull the 3 judge models (run once):
  ollama pull mistral
  ollama pull llama3.1
  ollama pull gemma2
  # 3. Ollama starts automatically — or manually: ollama serve
  

• gTTS — required only if you run generate_synthetic_cs.py --tts (Strategy B) Already included in requirements.txt. If missing, install manually:

  pip install gTTS
  

  Note: generate_synthetic_cs.py without --tts (Strategy A, text-only) does not need gTTS.

---

Full Run Order — Step by Step

Phase 1 — Data Collection & Preparation

cd dataset_pipeline

# Step 1a: Download from YouTube (dialect-name search queries)
python 01_download_audio.py

# Step 1b: Download from Facebook / other sites (paste URLs in script first)
python 01b_download_extra_urls.py

# Convert any .webm / .m4a files to .wav
python convert_webm.py

# Step 2: VAD segmentation → 5–15 second clips
python 02_segment_audio.py

# Step 3: Whisper transcription + CS detection (GPU strongly recommended)
python 03_auto_transcribe.py

# Step 4b: Auto-accept HIGH confidence transcripts (no_speech_prob < 0.30)
python 04b_auto_annotate.py

# Step 4c: LLM ensemble judges BORDERLINE clips (0.30 ≤ no_speech_prob < 0.60)
#          3 models vote: Mistral, LLaMA 3.1, Gemma 2 — majority wins
#          Requires Ollama running with all 3 models pulled (see Installation)
python 04c_llm_transcript_judge.py

# Step 5: Build source-isolated train/dev/test dataset splits
python 05_build_dataset.py

Output after Phase 1:

dataset_pipeline/
  raw_audio/          ← downloaded audio files
  segments/           ← 5–15 sec clips + segments_manifest.json
  transcripts/        ← transcripts.json, for_human_review.tsv, code_switched_clips.tsv
  final_dataset/
    train/manifest.csv + audio/
    dev/manifest.csv  + audio/
    test/manifest.csv + audio/

Data ASR Worker Merge Status

The four dataset_pipeline/Data ASR/transcripts_worker* folders can be merged with:

python dataset_pipeline/07_merge_data_asr_workers.py

The merged output is written to:

dataset_pipeline/Data ASR/transcripts_merged_all/

Merged files:

transcripts.json
for_human_review.tsv
code_switched.tsv
code_switched_clips.tsv

Current merge summary:

Source	Accepted clips
transcripts_worker1	12,524
transcripts_worker2	12,411
transcripts_worker3	12,444
transcripts_worker4	11,393
Merged total	48,772

If the expected total clip count is 52,016, then:

Metric	Count
Expected clips	52,016
Accepted merged clips	48,772
Missing/rejected clips	3,244
Explicitly logged skipped clips	242
Unlogged missing/rejected clips	3,002
Duplicate clip_ids across workers	0
Code-switched clips in merged output	2,056

Dialect distribution in the merged output:

Dialect	Clips
Sylhet	9,231
Sandwip	5,483
Old_Dhaka	5,460
Barishal	3,827
Chittagong	3,483
Khulna	3,310
Kishoreganj	2,741
Narail	2,601
Rangpur	2,387
Comilla	2,384
Narsingdi	1,972
Mymensingh	1,699
Habiganj	1,609
Tangail	1,321
Noakhali	1,264
Total	48,772

---

Phase 2 — Fine-Tuning

cd ../fine_tuning
pip install -r requirements.txt

# Prepare HuggingFace datasets (all 3 subsets at once — run once only)
python 01_prepare_hf_dataset.py

# Fine-tune Whisper large-v3 (3 experiments: all / bn_only / cs_only)
# Each run saves to fine_tuning/models/ft_whisper_{subset}/
python 02_finetune_whisper.py

# Fine-tune wav2vec2-XLS-R and WavLM (CTC models)
# Saves to fine_tuning/models/ft_wav2vec2_all/ and ft_wavlm_all/
python 03_finetune_wav2vec2.py

What each fine-tuning experiment produces:

Model saved to	Training data	Purpose
models/ft_whisper_all/	All clips (BN_ONLY + CS)	Proposed system
models/ft_whisper_bn_only/	Dialect-only clips	Ablation A
models/ft_whisper_cs_only/	CS clips only	Ablation B
models/ft_wav2vec2_all/	All clips	CTC baseline
models/ft_wavlm_all/	All clips	CTC baseline

GPU time estimates (A100 80GB):

• Whisper large-v3: ~8–12 hours per experiment
• wav2vec2 / WavLM: ~4–6 hours

---

Phase 3 — Evaluation

cd ../evaluation
pip install -r requirements.txt

# Zero-shot baselines (Whisper, Tugstugi/Ben-10, MMS, wav2vec2)
python 01_run_baselines.py

# Fine-tuned model predictions
python 02_run_finetuned.py

# LM shallow fusion for CTC models (run --tune_alpha for best results)
python 09_lm_shallow_fusion.py --tune_alpha

# Whisper size comparison (tiny → large-v3)
python 10_whisper_size_comparison.py

# Compute all metrics: WER, CER, MIX-ER, WER-BN, WER-EN, CS-F1
python 03_compute_metrics.py

# Build ablation table → ablation_table.tex (paste into LaTeX paper)
python 04_ablation_summary.py

# Generate 5 publication-ready plots
python 05_plot_results.py

# Statistical significance: bootstrap CI, p-values, Cohen's d
python 06_statistical_analysis.py

# Cross-dialect WER breakdown + leave-one-out experiment
python 07_cross_dialect_eval.py

# Error analysis: Sub/Del/Ins + English word survival
python 08_error_analysis.py

Output after Phase 3:

evaluation/
  results/
    *_predictions.json          ← raw predictions per model
    *_metrics.json              ← metrics per model
    all_metrics_summary.json    ← everything in one file
    ablation_table.tex          ← paste into paper
    ablation_table.csv
    statistical_report.txt
    cross_dialect/
    error_analysis/
  plots/
    wer_comparison.pdf/png
    dialect_wer_heatmap.pdf/png
    cs_f1_comparison.pdf/png
    cs_vs_bn_wer.pdf/png
    dialect_distribution.pdf/png
    whisper_size_comparison.pdf/png
    error_analysis.pdf/png

---

Phase 4 — Paper

cd ../asr_paper_draft

# Copy plots
cp ../evaluation/plots/*.png figures/

# Compile (requires MiKTeX or TeX Live)
make
# or manually:
pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex

After running all experiments, replace every \TODO{X.XX} in sections/06_results.tex and sections/07_analysis.tex with numbers from evaluation/results/all_metrics_summary.json.

---

Models & Metrics

Models evaluated

Model	Type	Zero-shot	Fine-tuned
Whisper large-v3	Seq2Seq	✓	✓ (3 variants)
Whisper tiny/base/small/medium	Seq2Seq	✓	✓
Tugstugi/whisper-bn (Ben-10)	Seq2Seq	✓	—
Meta MMS-1B	CTC	✓	—
wav2vec2-XLS-R	CTC	✓	✓ + LM
WavLM-Large	CTC	—	✓ + LM

Metrics computed

Metric	What it measures
WER	Overall word error rate
CER	Character error rate (important for Bengali morphology)
WER-BN	WER on Bengali-script words only
WER-EN	WER on English/Latin words only
MIX-ER	(WER-BN + WER-EN) / 2 on CS clips — CS-aware metric
CS-F1	F1 for detecting whether a clip is code-switched
EN-Survival	Fraction of English reference words preserved in hypothesis
Bootstrap CI	95% confidence interval for each model's WER
p-value	Paired bootstrap significance test (proposed vs each baseline)
Cohen's d	Effect size

---

Dataset Statistics

After segmentation (Phase 1, Step 2)

Dialect	Region	Clips	Hours
Sylhet	Northeast Bangladesh + UK diaspora	9,665	31.85
Old_Dhaka	Historic Dhaka city	5,767	17.84
Sandwip	Island, Chittagong coast	5,744	16.31
Barishal	South-central Bangladesh	4,048	13.57
Khulna	Southwest Bangladesh	3,540	11.40
Chittagong	Southeast Bangladesh	3,661	11.16
Kishoreganj	Central Bangladesh	3,038	9.52
Comilla	Southeast Bangladesh	2,724	9.39
Narail	Southwest Bangladesh	2,714	8.93
Rangpur	Northwest Bangladesh	2,650	8.67
Narsingdi	Central Bangladesh	2,150	6.63
Mymensingh	North-central Bangladesh	1,831	5.50
Habiganj	Northeast Bangladesh	1,694	5.28
Tangail	Central Bangladesh	1,396	4.77
Noakhali	Southeast Bangladesh	1,235	3.93
Total	15 dialects	51,857	164.74

Final Curated Dataset (After Augmentation & Splits)

Split	Clips	Hours	CS Clips	CS %
Train	78,122	247.38	6,662	8.5%
Dev	2,768	8.89	185	6.7%
Test	5,163	17.18	236	4.6%
Total	86,053	273.46	7,083	8.2%

(Note: The Train split includes augmented audio to balance underrepresented dialects and code-switched data. The Dev and Test sets remain strictly 100% real, unaugmented human speech. Roughly 10,000 original clips were safely filtered out during Phase 1 due to severe background noise or alignment errors).

---

Key Design Decisions

Why dialect-name queries, not region-name queries

Bad: "চট্টগ্রামের vlog english" → finds anyone from Chittagong, they speak standard Bengali for YouTube audience.

Good: "চাটগাঁইয়া ভাষায় comedy english" → only surfaces creators who consciously use the dialect — a strong signal for genuine dialectal content.

Why comedy/roast content

Comedy forces natural unscripted dialect. A creator cannot maintain a fake accent through jokes. This is the highest-precision query strategy for dialectal CS.

Why silver-standard transcription (no human annotators)

Whisper large-v3 with no_speech_prob < 0.30 produces transcripts with ~3–8% WER on clear speech. CS detection (Bengali vs. English words) is near-perfect because the scripts are visually distinct Unicode ranges. We validate on a 200-clip manual sample and report this WER in the paper.

Why Whisper over CTC for the proposed system

Whisper's seq2seq tokeniser contains both Bengali Unicode and Latin ASCII in one shared vocabulary. It produces mixed-script output (আমি phone কিনছি) in a single forward pass. CTC models need an explicit character vocabulary covering both scripts and cannot produce mixed-script output as naturally.

---

Requirements Summary

torch>=2.3.0              # GPU training
transformers>=4.40.0      # Whisper, wav2vec2, WavLM, MMS, HuBERT
datasets>=2.19.0          # HuggingFace dataset format
evaluate>=0.4.2           # WER metric
accelerate>=0.30.0        # Multi-GPU / mixed precision
openai-whisper>=20240930  # Zero-shot Whisper baseline
jiwer>=3.0.3              # WER / CER computation
pydub>=0.25.1             # Audio loading
webrtcvad-wheels>=2.0.10  # VAD segmentation (Python 3.12)
yt-dlp>=2024.11.18        # YouTube download
pandas>=2.2.0             # Dataset manifests
matplotlib>=3.9.0         # Plots
seaborn>=0.13.2           # Heatmaps
scipy>=1.13.0             # Statistical tests
imageio-ffmpeg>=0.5.1     # Bundled ffmpeg binary
pyctcdecode               # CTC beam search with LM (optional)
kenlm                     # n-gram LM (optional, for Step 09)

---