voxtral-mini-3b-tr-lora

Model Details

• Model name: voxtral-mini-3b-tr-lora
• Developed by: Yağmur Tuncer
• Shared by: Yağmur Tuncer
• Model type: Automatic Speech Recognition (ASR) – LoRA Adapter
• Language(s): Turkish (tr)
• License: Apache 2.0 (inherits base model license)
• Finetuned from: mistralai/Voxtral-Mini-3B-2507

---

Model Description

This model is a Turkish Automatic Speech Recognition (ASR) LoRA adapter fine-tuned on top of Voxtral-Mini-3B-2507 using QLoRA.

The base Voxtral model is kept fully frozen, and only a small set of Low-Rank Adaptation (LoRA) parameters are trained.
This results in a parameter-efficient, lightweight, and easily deployable ASR model specialized for Turkish speech.

Intended Use Cases

• Turkish speech transcription
• Domain adaptation experiments
• Research on parameter-efficient fine-tuning for multimodal ASR models

---

Model Sources

• Base Model: mistralai/Voxtral-Mini-3B-2507
• Repository: https://huggingface.co/y0mur/voxtral-mini-3b-tr-lora
• LoRA Method: Hu et al., 2021
• Paper: https://arxiv.org/abs/2106.09685

---

Uses

Direct Use

• Transcribing Turkish speech audio (16 kHz WAV recommended)
• ASR research using LoRA / PEFT
• Continuing fine-tuning from this LoRA adapter

Downstream Use

• Domain-specific ASR adaptation
• Speech analytics pipelines
• Dataset labeling / transcription automation

Out-of-Scope Use

• Languages other than Turkish
• Speaker identification
• Real-time streaming ASR without additional optimization

---

Bias, Risks, and Limitations

• The model inherits biases from the training datasets (Common Voice, FLEURS).
• Performance may degrade on:
  • Heavy accents
  • Code-switched speech
  • Noisy or low-quality audio
• The model has not been evaluated for demographic fairness.

---

Recommendations

Users are advised to:

• Validate performance on their own domain-specific data
• Avoid use in legal or safety-critical transcription without human review
• Experiment with decoding strategies (beam search, temperature tuning)

---

How to Get Started

import torch
from transformers import AutoProcessor, VoxtralForConditionalGeneration
from peft import PeftModel

BASE_MODEL = "mistralai/Voxtral-Mini-3B-2507"
LORA_REPO = "y0mur/voxtral-mini-3b-tr-lora"

processor = AutoProcessor.from_pretrained(BASE_MODEL)

base_model = VoxtralForConditionalGeneration.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype=torch.float16,
)

model = PeftModel.from_pretrained(base_model, LORA_REPO)
model.eval()