whisper-large-v3-chileno-lora

whisper-large-v3-chileno-lora

A LoRA adapter that helps Whisper Large-v3 transcribe Chilean Spanish from real call center phone calls, with spontaneous speech, slang and line noise. Made by HablaIP SpA.

Made by hablaip on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a LoRA adapter for openai/whisper-large-v3, tuned to transcribe Spanish as spoken in Chile on real call center phone calls. Those calls have spontaneous speech, local slang and phone line noise. HablaIP SpA made it within the CORFO project 25CH2-306482.

Who it suits: It suits teams that need text from Chilean Spanish phone audio, and developers who want to build that into a speech pipeline.

Industry | Who uses it | How they use it | Example
Call centers | Operations and quality teams in Chile | Turn recorded calls into text | Run the adapter over a call recording and read the transcript
Speech technology | Developers of speech recognition pipelines | Add the adapter on top of Whisper Large-v3 | Load it through PEFT with the base model
Speech research | Researchers of domain adaptation | Study a small LoRA on a narrow domain | Compare the 21.18% error rate with the untuned 34.76%

If your audio is read speech, another Spanish accent, or a different kind of call, the card has not measured it, so you may want to test first.

Where to use it:

• Python scripts with torch, transformers and peft
• Pipelines that feed 16 kHz audio arrays into Whisper
• Windows, macOS or Linux with Python

Use case | Who | What they get back
Call transcription | Call center teams | A literal Spanish transcript of a phone call
Baseline for fine-tuning | Researchers | A tuned model to compare against plain Whisper Large-v3
Speech analytics input | Developers | Text with filler words kept, ready for analysis

How to set it up: What you need first:

• Python with torch, transformers and peft
• The base model openai/whisper-large-v3, which the code loads for you
• Audio as a numpy array sampled at 16 kHz

1. Install the Python libraries the code imports.
2. Optionally fetch the adapter files with the Hugging Face tools.
3. Load the base model, then the adapter on top of it, as in the code below.

pip install -U huggingface_hub
hf download hablaip/whisper-large-v3-chileno-lora

How to use it: The code loads Whisper Large-v3, fixes the language to Spanish and the task to transcribe, applies the adapter, and decodes your audio. The audio variable stands for your own numpy array.

import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

processor = WhisperProcessor.from_pretrained(
    "openai/whisper-large-v3", language="spanish", task="transcribe"
)
model = WhisperForConditionalGeneration.from_pretrained(
    "openai/whisper-large-v3", torch_dtype=torch.bfloat16
)
model.generation_config.language = "spanish"
model.generation_config.task = "transcribe"
model.generation_config.forced_decoder_ids = None

model = PeftModel.from_pretrained(model, "hablaip/whisper-large-v3-chileno-lora")
model = model.eval()

# audio: arreglo numpy a 16kHz
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
    generated_ids = model.generate(input_features=inputs.input_features.to(torch.bfloat16), max_new_tokens=225)
texto = processor.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Tool | What it does
WhisperProcessor | Prepares the audio features and decodes the generated tokens into text
WhisperForConditionalGeneration | Loads the Whisper Large-v3 base model
PeftModel | Applies the LoRA adapter to the base model

Example: You pass in a 16 kHz recording of a Chilean phone call. The model returns the Spanish words as they were said, filler words and repetitions included, with no editorial cleanup.

For newcomers and experts: Think of it as an add-on for Whisper, the speech-to-text model, that hears Chilean call center Spanish better. A good first try is one short recording, run through the code above.

For experts, the LoRA uses rank 32, alpha 64 and dropout 0.05 on the q_proj and v_proj projections of the attention blocks. Roughly 1% of the parameters train and the base model stays frozen. The data is about 1,800 clips, around 8 hours, transcribed by hand word for word with no grammar fixes. The authors split data by whole call, so fragments of one conversation never land in both training and test.

Start from Whisper Large-v3
Freeze it and train small LoRA layers
Train on hand-transcribed call center clips
Load the adapter on top of the base model

Good to know:

Title: Word error rate on 153 call audios
Whisper Large-v3 untuned: 34.76%
This adapter: 21.18%

• Result: the card reports a 39% relative drop in errors on 78 distinct calls, with a bootstrap interval that does not cross the baseline.
• Domain: it is tuned for Chilean phone speech from call centers, and other domains were not characterized.
• Data: the training set is small and private, so it is not published with the model.
• Output: transcripts are verbatim, with filler words and repetitions, not cleaned for editing.
• License: apache-2.0
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads the base model and adapter from Hugging Face, and the code processes the audio you give it.
• Not for: clean editorial transcripts, or accents and speech types the card did not measure.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 79
• Likes: 6
• License: apache-2.0
• Languages: es
• Based on: openai/whisper-large-v3
• Tags: peft, safetensors, whisper, automatic-speech-recognition, spanish, chilean-spanish, call-center, lora

Links

• Model: https://huggingface.co/hablaip/whisper-large-v3-chileno-lora
• Files: https://huggingface.co/hablaip/whisper-large-v3-chileno-lora/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hablaip/whisper-large-v3-chileno-lora

Then follow the usage in its model card. Nothing in this kit runs by itself.