Keyword-Spotting

Model Card for hf-kws (Wav2Vec2 Keyword Spotting)

<!-- Provide a quick summary of what the model is/does. -->

A compact, end‑to‑end pipeline for training, evaluating, and deploying a Wav2Vec2‑based keyword spotting (KWS) model on Google Speech Commands v2. The repository includes offline and real‑time streaming inference, ONNX export, and AWS SageMaker deployment scripts.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

This project fine‑tunes a Wav2Vec2 audio classifier (e.g., facebook/wav2vec2-base) for keyword spotting on Speech Commands v2 using Hugging Face transformers/datasets. It supports microphone streaming with sliding‑window smoothing, file‑based inference, saved JSON metrics/plots, and a minimal SageMaker stack (train, realtime/serverless deploy, batch transform).

• Developed by: Amirhossein Yousefiramandi (GitHub: @amirhossein-yousefi)
• Model type: Audio Classification (Keyword Spotting) — Wav2Vec2 backbone with classification head
• Language(s) (NLP): English
• License: No explicit repository license file found (verify before redistribution)
• Finetuned from model : facebook/wav2vec2-base (16 kHz)

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: https://github.com/amirhossein-yousefi/keyword-spotting
• Paper : Warden, P. (2018). Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition. arXiv:1804.03209.

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

• On‑device or edge keyword detection for small command sets (e.g., “yes/no/up/down/stop/go”).
• Real‑time wake word / trigger prototypes via the included streaming inference script.
• Batch scoring of short audio clips for command presence via CLI or SageMaker Batch Transform.

Downstream Use

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

• Fine‑tune on custom keyword lists or languages (swap dataset, keep pipeline).
• Distillation/quantization for mobile deployment (roadmap mentions TFLite/CoreML).

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

• Open‑vocabulary ASR or general transcription.
• Long‑form audio or multi‑speaker diarization.
• Safety‑critical activation (e.g., medical/industrial controls) without rigorous evaluation and fail‑safes.
• Always‑on surveillance scenarios without clear user consent and privacy controls.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

• Language & domain bias: Trained on English, one‑second command words—limited transfer to other languages, accents, far‑field mics, or noisy environments without adaptation.
• Vocabulary constraints: Detects from a fixed label set; out‑of‑vocabulary words may map to “unknown” or be misclassified.
• Data licensing: Ensure CC‑BY‑4.0 attribution when redistributing models trained on Speech Commands.

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Evaluate on target devices/microphones; add noise augmentation and tune detection thresholds for deployment context.

Usage in HuggingFace (Recommended)

from transformers import AutoFeatureExtractor, AutoModelForAudioClassification, pipeline

model_id = "Amirhossein75/Keyword-Spotting"
# Option A — simple:
clf = pipeline("audio-classification", model=model_id)
print(clf("path/to/1sec_16kHz.wav"))

# Option B — manual pre/post:
fe = AutoFeatureExtractor.from_pretrained(model_id)
model = AutoModelForAudioClassification.from_pretrained(model_id)

import soundfile as sf, torch
wave, sr = sf.read("path/to/1sec_16kHz.wav")
inputs = fe(wave, sampling_rate=sr, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
pred_id = int(logits.argmax(-1))
print(model.config.id2label[pred_id])

Note : it is better not to use the AutoProcessor

How to Get Started with the Model

Use the code below to get started with the model.

# clone and install
git clone https://github.com/amirhossein-yousefi/keyword-spotting
cd keyword-spotting
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install --upgrade pip
pip install -r requirements.txt

# train (example)
python -m src.train   --checkpoint facebook/wav2vec2-base   --output_dir ./checkpoints/kws_w2v2   --num_train_epochs 8   --per_device_train_batch_size 16   --per_device_eval_batch_size 16

# single-file inference
python -m src.infer   --model_dir ./checkpoints/kws_w2v2   --wav_path /path/to/your.wav   --top_k 5

# streaming (microphone)
python -m src.stream_infer --model_dir ./checkpoints/kws_w2v2

# evaluate
python -m src.evaluate_fn --model_dir ./checkpoints/kws_w2v2

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

• Dataset: Google Speech Commands v2 (1‑second WAVs, 16 kHz; English; CC‑BY‑4.0). Typical label set includes “yes/no, up/down, left/right, on/off, stop/go,” plus auxiliary words and silence/unknown classes.

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Preprocessing

• Resampled/processed at 16 kHz.
• Augmentations: time‑shift, noise, random gain.

Training Hyperparameters

• Training regime: fp32 (example; adjust as needed)
• Backbone: facebook/wav2vec2-base (audio classification head).
• Epochs (example): 8
• Batch size (example): 16 train / 16 eval
• Framework: PyTorch + Hugging Face transformers

Speeds, Sizes, Times

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

• Example environment: Single NVIDIA GeForce RTX 3080 Ti Laptop GPU (16 GB), PyTorch 2.8.0+cu129, CUDA 12.9.
• Reported training runtime: ~3,446.3 s for the default run (see repository logs/README).

Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

Testing Data, Factors & Metrics

Testing Data

• Speech Commands v2 test split.

Factors

• Evaluate by keyword, speaker, noise type/level, and device/mic to assess robustness.

Metrics

• Accuracy, F1, Precision, Recall, and Cross‑entropy loss; plus runtime and throughput.

Results

Below are the aggregated metrics at epoch 10.

Summary

The pipeline reproduces standard Wav2Vec2 KWS performance on Speech Commands; tailor thresholds and augmentations for deployment.

Model Examination

<!-- Relevant interpretability work for the model goes here -->

• Inspect per‑class confusion matrices and score distributions from saved metrics to identify false‑positive/negative patterns.

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: Single NVIDIA GeForce RTX 3080 Ti Laptop GPU
• Hours used: ~0.96 h (example run)

Technical Specifications

Model Architecture and Objective

• Architecture: Wav2Vec2 (self‑supervised acoustic encoder) + classification head for KWS.

Compute Infrastructure

Hardware

• Example: NVIDIA RTX 3080 Ti Laptop, 16 GB VRAM.

Software

• PyTorch 2.8.0+cu129, CUDA driver 12.9; Hugging Face transformers/datasets.

Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

BibTeX (Dataset):

@article{warden2018speechcommands,
  author = {Warden, Pete},
  title  = {Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition},
  journal= {arXiv e-prints},
  eprint = {1804.03209},
  year   = {2018},
  month  = apr,
  url    = {https://arxiv.org/abs/1804.03209}
}

APA (Dataset):
Warden, P. (2018). Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition. arXiv:1804.03209.

Glossary

<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->

• KWS: Keyword Spotting — detecting a small set of pre‑registered words in short audio clips.
• Streaming inference: Frame‑by‑frame scoring with smoothing over a sliding window.

More Information

• Speech Commands dataset card: https://huggingface.co/datasets/google/speech_commands
• Wav2Vec2 model docs: https://huggingface.co/docs/transformers/en/model_doc/wav2vec2

Model Card Authors

• Amirhossein Yousefiramandi

Model Card Contact

Please open a GitHub Issue in this repository with questions or requests.