whisper-algerian-darja-small API & Inference Endpoint

Overview

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

Built on top of OpenAI's Whisper Small (openai/whisper-small, 244M parameters), this model incorporates parameter-efficient LoRA adapters trained with 4-bit quantization (QLoRA) over a sequential 3-phase curriculum covering conversational podcasts, spontaneous storytelling, and cultural narratives from the OddAdmix Algerian speech collection.

Key Highlights

Native Algerian Dialect Adaptation: Handles authentic Darja phonetics, vocabulary, morphology, and spontaneous conversational speech patterns.
Parameter-Efficient LoRA (PEFT): Trained on 25.95M parameters (9.70% of total model weights), allowing lightweight adapter storage and fast inference.
Significant Accuracy Gains: Achieved 14.87% WER on storytelling evaluation subsets and 27.54% WER on raw narrative benchmarks.
Sequential Streaming Curriculum: Trained end-to-end via multi-phase streaming on Hugging Face CDN without disk bottlenecking.

Evaluation & Benchmark Results

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

Table with columns: Phase, Domain / Dataset, Epochs / Steps, Learning Rate, Best WER (%), Final Eval Loss
Phase	Domain / Dataset	Epochs / Steps	Learning Rate	Best WER (%)	Final Eval Loss
Phase 1	Kahwa Podcast (`oddadmix/arabic-audio-collection-algerian-kahwa-postcast`)	2 epochs (4,942 steps)	$1 \times 1 0 - 4$	34.85%	0.521
Phase 2	Loubna Stories (`oddadmix/arabic-audio-collection-algerian-loubna-stories`)	2 epochs (10,324 steps)	$5 \times 1 0 - 5$	14.87%	0.312
Phase 3	Rawi Storytelling (`oddadmix/arabic-audio-collection-algerian-rawi`)	1 epoch (562 steps)	$2 \times 1 0 - 5$	27.54%	0.2548

Cumulative Training Progress: Total training ran for 15,829 cumulative optimization steps with cosine annealing learning rate schedules and automatic best-adapter preservation per phase.

Architecture & Training Methodology

text

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

Model Configuration

Base Architecture: openai/whisper-small (Encoder-Decoder Transformer)
Base Model Parameters: 241,734,912
Trainable LoRA Parameters: 25,952,256 (9.695%)
Total Parameters: 267,687,168
LoRA Hyperparameters:
- Rank ( $r$ ): 64
- Scaling factor ( $α$ ): 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

Audio Normalization: Resampled to mono 16,000 Hz float32 arrays on-the-fly.
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 $\to$ ا
- Yaa / Alef Maksura normalization: ى $\to$ ي
- Punctuation stripping (standard Latin & Arabic punctuation ،, ؛, ؟, «, »).
Audio Quality Filtering:
- Duration bounds: $0.5 s \leq duration \leq 30.0 s$
- Character density filter: $1.0 \leq duration len (transcript) \leq 25.0 chars/sec$

Quickstart & Inference

1. Installation

bash

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

2. High-Level Usage with `transformers.pipeline`

python

import torchfrom transformers import pipeline # Initialize speech recognition pipeline with PEFT adapterpipe = pipeline(    task="automatic-speech-recognition",    model="touati-kamel/whisper-algerian-darja-small",    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

python

import torchimport librosafrom transformers import WhisperProcessor, WhisperForConditionalGenerationfrom peft import PeftModel device = "cuda" if torch.cuda.is_available() else "cpu"model_id = "openai/whisper-small"adapter_id = "touati-kamel/whisper-algerian-darja-small" # 1. Load Processorprocessor = WhisperProcessor.from_pretrained(model_id, language="arabic", task="transcribe") # 2. Load Base Model and Apply LoRA Adapterbase_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 Audioaudio, sr = librosa.load("path/to/audio.mp3", sr=16000)input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_featuresif torch.cuda.is_available():    input_features = input_features.to("cuda", dtype=torch.float16) # 4. Generate Transcriptionforced_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 Outputtranscription = 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:

python

import reimport 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()

Training Hyperparameters & Setup

Table with columns: Hyperparameter, Value
Hyperparameter	Value
Base Model	`openai/whisper-small` (244M params)
Quantization	4-bit NF4 (`BitsAndBytesConfig`)
Hardware	1x NVIDIA Tesla T4 GPU (16 GB VRAM)
Per-Device Batch Size	8
Gradient Accumulation Steps	4 (Effective Batch Size = 32)
Mixed Precision	FP16 (`fp16=True`)
Gradient Checkpointing	Enabled
Optimizer	AdamW
Learning Rate Schedule	Cosine Annealing with Warmup
Warmup Steps	100 (Phase 1), 50 (Phase 2), 30 (Phase 3)
Evaluation Strategy	Every 300 steps with WER computation
Checkpoint Strategy	Automatic Best WER saving + Hugging Face Hub upload

Intended Uses & Limitations

Intended Uses

Speech-to-text transcription for Algerian podcasts, YouTube content, voice messages, and media.
Algerian Darija conversational assistants and voice search interfaces.
Transcription and subtitling for Algerian cultural and educational audio.

Limitations & Biases

Code-Switching: Algerian Darja frequently blends Arabic with French and Berber/Tamazight loanwords. While French tag handling was integrated during preprocessing, heavy French sentences may be transcribed phonetically into Arabic script.
Regional Dialectal Variations: The training data prominently covers Central and Western dialects. Eastern (Constantinois/Annaba) or Southern (Sahara) variants with distinct phonetic shifts may experience slight variance in accuracy.
Noisy Acoustic Environments: Performance is optimal on clean speech (podcasts, stories); background music or intense street noise may increase WER.

Datasets & Citations

Training Datasets

BibTeX Citation

bibtex

@misc{touati2026whisper_algerian_darja,  author = {Kamel Touati},  title = {Whisper Small 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-small}}}

bibtex

@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}}

Phase

Domain / Dataset

Epochs / Steps

Learning Rate

Best WER (%)

Final Eval Loss

Phase 1

Kahwa Podcast (oddadmix/arabic-audio-collection-algerian-kahwa-postcast)

2 epochs (4,942 steps)

1 \times 1 0 - 4

34.85%

0.521

Phase 2

Loubna Stories (oddadmix/arabic-audio-collection-algerian-loubna-stories)

2 epochs (10,324 steps)

5 \times 1 0 - 5

14.87%

0.312

Phase 3

Rawi Storytelling (oddadmix/arabic-audio-collection-algerian-rawi)

1 epoch (562 steps)

2 \times 1 0 - 5

27.54%

0.2548

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

import torchfrom transformers import pipeline # Initialize speech recognition pipeline with PEFT adapterpipe = pipeline( task="automatic-speech-recognition", model="touati-kamel/whisper-algerian-darja-small", 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"])

import torchimport librosafrom transformers import WhisperProcessor, WhisperForConditionalGenerationfrom peft import PeftModel device = "cuda" if torch.cuda.is_available() else "cpu"model_id = "openai/whisper-small"adapter_id = "touati-kamel/whisper-algerian-darja-small" # 1. Load Processorprocessor = WhisperProcessor.from_pretrained(model_id, language="arabic", task="transcribe") # 2. Load Base Model and Apply LoRA Adapterbase_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 Audioaudio, sr = librosa.load("path/to/audio.mp3", sr=16000)input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_featuresif torch.cuda.is_available(): input_features = input_features.to("cuda", dtype=torch.float16) # 4. Generate Transcriptionforced_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 Outputtranscription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]print("Algerian Darja Output:", transcription)

import reimport 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()

Hyperparameter

Value

Base Model

openai/whisper-small (244M params)

Quantization

4-bit NF4 (BitsAndBytesConfig)

Hardware

1x NVIDIA Tesla T4 GPU (16 GB VRAM)

Per-Device Batch Size

Gradient Accumulation Steps

4 (Effective Batch Size = 32)

Mixed Precision

FP16 (fp16=True)

Gradient Checkpointing

Enabled

Optimizer

AdamW

Learning Rate Schedule

Cosine Annealing with Warmup

Warmup Steps

100 (Phase 1), 50 (Phase 2), 30 (Phase 3)

Evaluation Strategy

Every 300 steps with WER computation

Checkpoint Strategy

Automatic Best WER saving + Hugging Face Hub upload

@misc{touati2026whisper_algerian_darja, author = {Kamel Touati}, title = {Whisper Small 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-small}}}

@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}}

whisper-algerian-darja-small

README

Overview

Key Highlights

Evaluation & Benchmark Results

Architecture & Training Methodology

Model Configuration

Data Preprocessing & Arabic Normalization

Quickstart & Inference

1. Installation

2. High-Level Usage with `transformers.pipeline`

3. Native PyTorch + `PeftModel` Inference

Complete Normalization Function (Recommended for Evaluation)

Training Hyperparameters & Setup

Intended Uses & Limitations

Intended Uses

Limitations & Biases

Datasets & Citations

Training Datasets

BibTeX Citation

Explore FriendliAI today

README

Overview

Key Highlights

Evaluation & Benchmark Results

Architecture & Training Methodology

Model Configuration

Data Preprocessing & Arabic Normalization

Quickstart & Inference

1. Installation

2. High-Level Usage with `transformers.pipeline`

3. Native PyTorch + `PeftModel` Inference

Complete Normalization Function (Recommended for Evaluation)

Training Hyperparameters & Setup

Intended Uses & Limitations

Intended Uses

Limitations & Biases

Datasets & Citations

Training Datasets

BibTeX Citation

whisper-algerian-darja-small

README

Overview

Key Highlights

Evaluation & Benchmark Results

Architecture & Training Methodology

Model Configuration

Data Preprocessing & Arabic Normalization

Quickstart & Inference

1. Installation

2. High-Level Usage with transformers.pipeline

3. Native PyTorch + PeftModel Inference

Complete Normalization Function (Recommended for Evaluation)

Training Hyperparameters & Setup

Intended Uses & Limitations

Intended Uses

Limitations & Biases

Datasets & Citations

Training Datasets

BibTeX Citation

Explore FriendliAI today

README

Overview

Key Highlights

Evaluation & Benchmark Results

Architecture & Training Methodology

Model Configuration

Data Preprocessing & Arabic Normalization

Quickstart & Inference

1. Installation

2. High-Level Usage with transformers.pipeline

3. Native PyTorch + PeftModel Inference

Complete Normalization Function (Recommended for Evaluation)

Training Hyperparameters & Setup

Intended Uses & Limitations

Intended Uses

Limitations & Biases

Datasets & Citations

Training Datasets

BibTeX Citation

2. High-Level Usage with `transformers.pipeline`

3. Native PyTorch + `PeftModel` Inference

2. High-Level Usage with `transformers.pipeline`

3. Native PyTorch + `PeftModel` Inference