---
library_name: transformers, peft, torch
tags: [asr, whisper, finetune, atc, aircraft, communications, english]
---

# Model Card for Model ID

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

[SUMMARY HERE]

## Model Details

### Model Description

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

- **Developed by:** Jesse Arzate
- **Model type:** Sequence-to-Sequence (Seq2Seq) Transformer-based model
- **Language(s) (NLP):** English
- **License:** [More Information Needed]
- **Finetuned from model [optional]:** Whisper ASR: distil-large-v3

### Model Sources [optional]

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

- **Repository:** https://github.com/Vaibhavs10/fast-whisper-finetuning

## 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. -->

[More Information Needed]

### Downstream Use [optional]

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

[More Information Needed]

### Out-of-Scope Use

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

[More Information Needed]

## Bias, Risks, and Limitations

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

[More Information Needed]

### 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. More information needed for further recommendations.

## How to Get Started with the Model

Use the code below to get started with the model. 
```python
from transformers import (
    AutomaticSpeechRecognitionPipeline,
    WhisperForConditionalGeneration,
    WhisperTokenizer,
    WhisperProcessor,
)
from peft import PeftModel, PeftConfig


peft_model_id = "baileyarzate/whisper-distil-large-v3-atc-english" # huggingface model path
language = "en"
task = "transcribe"
device = 'cuda'
peft_config = PeftConfig.from_pretrained(peft_model_id)
model = WhisperForConditionalGeneration.from_pretrained(
    peft_config.base_model_name_or_path, device_map="cuda"
).to(device)

model = PeftModel.from_pretrained(model, peft_model_id).to(device)
tokenizer = WhisperTokenizer.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
processor = WhisperProcessor.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
feature_extractor = processor.feature_extractor
forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task=task)
pipe = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor)
model.config.use_cache = True

def transcribe(audio):
    with torch.cuda.amp.autocast():
        text = pipe(audio, generate_kwargs={"forced_decoder_ids": forced_decoder_ids}, max_new_tokens=255)["text"]
    return text
    
transcriptions_finetuned = []
for i in tqdm(range(len(df_subset))):
    # When you only have audio file path
    #transcriptions_finetuned.append(transcribe(librosa.load(df["path"][i], sr = 16000, offset = df["start"][i], duration = df["stop"][i] - df["start"][i])[0])) #,model
    # When you have audio array, saves time
    transcriptions_finetuned.append(transcribe(df_subset['array'].iloc[i]))
transcriptions_finetuned = pd.DataFrame(transcriptions_finetuned, columns=['transcription_finetuned'])
df_subset = df_subset.reset_index().drop(columns=['index'])
df_subset = pd.concat([df_subset, transcriptions_finetuned], axis=1)
```

## 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: ATC audio recordings from actual flight operations.
Size: ~250 hours of annotated data.

### Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
Modeled the procedure after: https://github.com/Vaibhavs10/fast-whisper-finetuning

#### Preprocessing [optional]

Preprocessing: Striped leading and trailing whitespaces from transcript sentences. Removed any sentences containing the phrase "UNINTELLIGIBLE" to filter out unclear or garbled speech. Removed filler words such as "ah" or "uh". 


#### Training Hyperparameters

- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
```python 
training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2, 
    learning_rate=5e-4, 
    warmup_steps=100,
    num_train_epochs=3,
    fp16=True,
    per_device_eval_batch_size=4,
    generation_max_length=128,
    logging_steps=100,
    save_steps=500,
    save_total_limit=3,
    remove_unused_columns=False,  # required as the PeftModel forward doesn't have the signature of the wrapped model's forward
    label_names=["labels"],  # same reason as above
)
```
#### Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
Inference time is about 2 samples per second with an RTX A2000.


## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->
Final training loss: 0.103

### Testing Data, Factors & Metrics

#### Testing Data

<!-- This should link to a Dataset Card if possible. -->
Dataset: ATC audio recordings from actual flight operations.
Size: ~250 hours of annotated data.
Randomly sampled 20% of the data with seed = 42.

[More Information Needed]

#### Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

[More Information Needed]

#### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

Word Error Rate, Normalized Word Error Rate

### Results

Mean WER for 500 test samples: 0.145 with 95% confidence interval: (0.123, 0.167)

#### Summary

[IN PROGRESS]


## 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](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).

- **Hardware Type:** RTX A2000
- **Hours used:** 24
- **Cloud Provider:** Private Infrustructure
- **Compute Region:** Southern California
- **Carbon Emitted:** 1.57 kg

## Technical Specifications [optional]

### Model Architecture and Objective

[More Information Needed]

### Compute Infrastructure

[More Information Needed]

#### Hardware

- **CPU**: AMD EPYC 7313P 16-Core Processor 3.00 GHz
- **GPU**: NVIDIA RTX A2000
- **vRAM**: 6GB
- **RAM**: 128GB

#### Software

- **OS**: Windows 11 Enterprise - 21H2
- **Python**: Python 3.10.14

## Citation [optional]

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

**BibTeX:**

[More Information Needed]

**APA:**

[More Information Needed]

## Model Card Contact

Jesse Arzate: baileyarzate@gmail.com