rir-i/legacyagent-qwen3-asr-lora

LegacyAgent — Qwen3-ASR LoRA

A domain-adapted automatic speech recognition (ASR) model based on Qwen/Qwen3-ASR-1.7B-hf, fine-tuned with QLoRA on U.S. Supreme Court oral-argument audio.

The model focuses on improving transcription quality for legal and entity-heavy courtroom speech while keeping the number of trainable parameters small through parameter-efficient LoRA adaptation.

Model Details

Model Description

This model is a QLoRA fine-tuned version of Qwen/Qwen3-ASR-1.7B-hf for English courtroom speech.

The model was trained on U.S. Supreme Court oral-argument audio using a case-disjoint train/validation setup. The evaluation focuses both on overall transcription quality and entity-span recognition, where errors in names, organizations, legal terminology, and other important spans are particularly costly.

Model Details

• —Developed by: Riri
• —Funded by: Personal project
• —Shared by: Riri
• —Model type: Automatic Speech Recognition (ASR)
• —Language(s): English
• —License: See the base model license
• —Finetuned from model: Qwen/Qwen3-ASR-1.7B-hf

Model Sources

• —Repository: https://github.com/Rupinder51120/LegacyAgent
• —Paper: N/A
• —Demo: N/A

Uses

Direct Use

The adapter can be used with the base Qwen3-ASR model for automatic transcription of English courtroom and legal-domain speech.

It is primarily intended for research and experimentation with domain-adapted ASR.

Downstream Use

The model can be integrated into ASR pipelines for:

• —Legal and courtroom transcription
• —Speech-to-text research
• —Domain-specific ASR evaluation
• —Entity-aware transcription systems
• —Legal AI research workflows

Out-of-Scope Use

This model should not be used:

• —As a replacement for official court transcripts
• —As a source of authoritative legal records
• —For making legal decisions without human verification
• —In safety-critical applications without appropriate validation
• —As a general-purpose ASR system for domains substantially different from courtroom speech

Bias, Risks, and Limitations

ASR systems can produce transcription errors caused by accents, speaker variation, recording quality, background noise, domain-specific terminology, and uncommon names.

Errors in legal names, organizations, citations, numbers, and other entities can be particularly consequential.

Performance outside U.S. Supreme Court oral-argument speech may differ from the reported benchmark results.

The reported evaluation should therefore be interpreted as a domain-specific benchmark rather than a guarantee of performance on unseen domains.

Recommendations

Users should verify important transcriptions against the original audio, particularly when the output contains names, legal terminology, numbers, quotations, or other legally significant information.

The model should be evaluated on the target deployment domain before being used in production.

How to Get Started with the Model

Install the required libraries:

pip install transformers peft bitsandbytes