whisper-algerian-darja-medium

Whisper Medium — Algerian Arabic (Darja / الدارجة الجزائرية)

<p align="center"> <img src="https://raw.githubusercontent.com/huggingface/transformers/main/docs/source/en/imgs/whisper_architecture.png" alt="Whisper Architecture" width="700"/> </p> <p align="center"> <a href="https://huggingface.co/touati-kamel/whisper-algerian-darja-medium"><img src="https://img.shields.io/badge/Hugging%20Face-Model%20Card-orange?style=flat-square&logo=huggingface" alt="Hugging Face Model"></a> <a href="https://github.com/openai/whisper"><img src="https://img.shields.io/badge/Base%20Model-OpenAI%20Whisper--medium-blue?style=flat-square" alt="Base Model"></a> <a href="https://github.com/huggingface/peft"><img src="https://img.shields.io/badge/PEFT-LoRA%204--bit%20(QLoRA)-purple?style=flat-square" alt="PEFT LoRA"></a> <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-green?style=flat-square" alt="License"></a> <a href="https://wandb.ai/k_touati-estin/whisper-algerian-darja/runs/bypsqt51"><img src="https://img.shields.io/badge/W%26B-Training%20Logs-gold?style=flat-square&logo=weightsandbiases" alt="Weights and Biases"></a> </p>

---

Overview

touati-kamel/whisper-algerian-darja-medium is an Automatic Speech Recognition (ASR) model specifically fine-tuned for Algerian Arabic (Darja / الدارجة الجزائرية).

Built on top of OpenAI's Whisper Medium (openai/whisper-medium, 769M base parameters / 833M total parameters), this model incorporates parameter-efficient LoRA adapters trained with 4-bit quantization (QLoRA) over a comprehensive sequential 3-phase curriculum covering conversational podcasts, spontaneous storytelling, and cultural narratives from the OddAdmix Algerian speech collection.

Key Highlights

• Native Algerian Dialect Adaptation: Exceptional comprehension of authentic Algerian Darja vocabulary, morphology, fast colloquial speech, and code-mixed expressions.
• Superior Acoustic & Language Modeling: Leveraging the 24-layer medium Whisper architecture for vastly superior contextual modeling compared to smaller model variants.
• Parameter-Efficient LoRA (PEFT): Trained on 69.21M parameters (8.31% of total model weights), allowing compact adapter storage and fast inference while preserving Whisper's general acoustic features.
• Low Error Rates: Achieved 0.34% WER on storytelling evaluation subsets, 0.68% WER on conversational podcast subsets, and 0.95% WER on cultural narratives.
• Sequential Streaming Curriculum: Trained end-to-end for 31,661 cumulative optimization steps with dynamic zero-disk streaming on Hugging Face CDN.

---

Evaluation & Benchmark Results

The model was evaluated iteratively across the three domain splits using Word Error Rate (WER %) and Cross-Entropy Evaluation Loss with Arabic dialect text normalization:

Cumulative Training Progress: Total training ran for 31,661 cumulative optimization steps with cosine annealing learning rate schedules and automatic best-adapter preservation per phase.

Comparison: Whisper Small vs. Whisper Medium

---

Architecture & Training Methodology

┌────────────────────────────────────────────────────────┐
│                   OpenAI Whisper-Medium                │
│              (4-bit NF4 Quantization Base)             │
│                 (763.86M Frozen Params)                │
└──────────────────────────┬─────────────────────────────┘
                           │
             ┌─────────────┴─────────────┐
             │    LoRA Adapters (r=64)   │
             │   Target: q, k, v, out,   │
             │           fc1, fc2        │
             │    (69.21M Trainable)     │
             └─────────────┬─────────────┘
                           │
       ┌────────────────────┴────────────────────┐
       │     Sequential Curriculum Learning      │
       ├─────────────────────────────────────────┤
       │ Phase 1: Kahwa Podcast (Conversational) │
       │                    ▼                    │
       │ Phase 2: Loubna Stories (Expressive)    │
       │                    ▼                    │
       │ Phase 3: Rawi Narratives (Storytelling) │
       └─────────────────────────────────────────┘

Model Configuration

• Base Architecture: openai/whisper-medium (Encoder-Decoder Transformer, 24 encoder layers, 24 decoder layers, 16 attention heads)
• Base Model Parameters: 763,857,920
• Trainable LoRA Parameters: 69,206,016 (8.3074%)
• Total Parameters: 833,063,936
• LoRA Hyperparameters:
  • Rank ($r$): 64
  • Scaling factor ($\alpha$): 128
  • LoRA Dropout: 0.05
  • Target Modules: q_proj, k_proj, v_proj, out_proj, fc1, fc2
  • Bias: none
• Quantization: 4-bit NormalFloat4 (NF4) with FP16 compute dtype (bitsandbytes)

Data Preprocessing & Arabic Normalization

1. Audio Normalization: Resampled to mono 16,000 Hz float32 arrays on-the-fly.
2. Text Sanitation:
   • Removal of French annotations/tags: [French: ...], [FR: ...]
   • Stripping non-speech markers and bracketed tokens: [...], <...>
   • Arabic Diacritics (Harakat / Tashkeel) removal: \u064B to \u0652, \u0670
   • Tatweel (Kashida) removal: \u0640
   • Alef normalization: إ, أ, آ $\rightarrow$ ا
   • Yaa / Alef Maksura normalization: ى $\rightarrow$ ي
   • Punctuation stripping (standard Latin & Arabic punctuation ،, ؛, ؟, «, »).
3. Audio Quality Filtering:
   • Duration bounds: $0.5\text{s} \le \text{duration} \le 30.0\text{s}$
   • Character density filter: $1.0 \le \frac{\text{len}(\text{transcript})}{\text{duration}} \le 25.0\text{ chars/sec}$

---

Quickstart & Inference

1. Installation

pip install --upgrade transformers peft torch torchaudio soundfile librosa jiwer bitsandbytes accelerate

2. High-Level Usage with transformers.pipeline

import torch
from transformers import pipeline

# Initialize speech recognition pipeline with PEFT adapter
pipe = pipeline(
    task="automatic-speech-recognition",
    model="touati-kamel/whisper-algerian-darja-medium",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device=0 if torch.cuda.is_available() else "cpu",
    chunk_length_s=30,
)

# Transcribe an audio file (16kHz WAV or MP3)
result = pipe(
    "path/to/algerian_audio.mp3",
    generate_kwargs={"language": "arabic", "task": "transcribe"}
)

print("Transcription (Darja):", result["text"])

3. Native PyTorch + PeftModel Inference

import torch
import librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "openai/whisper-medium"
adapter_id = "touati-kamel/whisper-algerian-darja-medium"

# 1. Load Processor
processor = WhisperProcessor.from_pretrained(model_id, language="arabic", task="transcribe")

# 2. Load Base Model and Apply LoRA Adapter
base_model = WhisperForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto" if torch.cuda.is_available() else None,
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

# 3. Load and Preprocess Audio
audio, sr = librosa.load("path/to/audio.mp3", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
if torch.cuda.is_available():
    input_features = input_features.to("cuda", dtype=torch.float16)

# 4. Generate Transcription
forced_decoder_ids = processor.get_decoder_prompt_ids(language="arabic", task="transcribe")
with torch.no_grad():
    predicted_ids = model.generate(
        input_features,
        forced_decoder_ids=forced_decoder_ids,
        max_new_tokens=225
    )

# 5. Decode Output
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print("Algerian Darja Output:", transcription)

---

Complete Normalization Function (Recommended for Evaluation)

To align with the training evaluation standard, use the following text normalizer:

import re
import string

_ARABIC_DIACRITICS = "\u064B\u064C\u064D\u064E\u064F\u0650\u0651\u0652\u0670"
_TATWEEL = "\u0640"
_PUNCT_MAP = {ord(c): None for c in string.punctuation + "\u060C\u061B\u061F\u00AB\u00BB"}

def normalize_darja_text(text: str) -> str:
    if not text:
        return ""
    # Strip diacritics & tatweel
    text = text.translate({ord(c): None for c in _ARABIC_DIACRITICS})
    text = text.replace(_TATWEEL, "")
    # Normalize Alef and Yaa
    text = text.replace("\u0625", "\u0627").replace("\u0623", "\u0627").replace("\u0622", "\u0627")
    text = text.replace("\u0649", "\u064A")
    # Remove punctuation
    text = text.translate(_PUNCT_MAP)
    # Collapse whitespace
    return " ".join(text.split()).strip()

---

Evaluation Benchmarks

1. Casablanca Algerian Benchmark

Evaluates Zero-Shot foundation Whisper vs. Fine-Tuned models on UBC-NLP/Casablanca (Algeria test split, 843 authentic dialectal utterances):

2. YODAS v3 Algerian Dialect Split (dz)

Evaluates Zero-Shot foundation Whisper vs. Fine-Tuned models on the Algerian partition of oddadmix/yodas-v3-ar-labeled-split (split dz, 3,000 utterances / 14.31 audio hours across 1,552 distinct video recordings):

---

Training Hyperparameters & Setup

---

Intended Uses & Limitations

Intended Uses

• High-accuracy speech-to-text transcription for Algerian podcasts, YouTube content, interviews, and media.
• Algerian Darija voice assistants, customer service transcription, and interactive voice response (IVR).
• Archival transcription and subtitling for Algerian cultural heritage, educational audio, and spoken narratives.

Limitations & Biases

• Code-Switching: Algerian Darja frequently blends Arabic with French and Berber/Tamazight loanwords. While French tag handling was integrated during preprocessing, intense mixed French sentences may occasionally be transcribed phonetically into Arabic script.
• Regional Dialectal Variations: The training data prominently covers Central (Algiers) and Western (Oran/Mostaganem) dialects. Eastern or Saharan accents with distinct phonetic nuances may exhibit slightly higher variance.
• Extreme Background Noise: Best results are achieved on clear speech. Extremely noisy field recordings or heavy background music may affect transcription accuracy.

---

Datasets & Citations

Training Datasets

• oddadmix/arabic-audio-collection-algerian-kahwa-postcast
• oddadmix/arabic-audio-collection-algerian-loubna-stories
• oddadmix/arabic-audio-collection-algerian-rawi

BibTeX Citation

@misc{touati2026whisper_algerian_darja_medium,
  author = {Kamel Touati},
  title = {Whisper Medium Fine-Tuned for Algerian Arabic (Darja)},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/touati-kamel/whisper-algerian-darja-medium}}
}

@article{radford2022whisper,
  title={Robust Speech Recognition via Large-Scale Weak Supervision},
  author={Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  journal={arXiv preprint arXiv:2212.04356},
  year={2022}
}