---
language:
- en
tags:
- speaker-profiling
- speaker-language-model
- speech
- zero-shot
- descriptive-profiling
- colmbo
- custom_code
license: apache-2.0
datasets:
- cmu-mlsp/TEARS
arxiv: 2506.09375
---
# ๐ต CoLMbo
### **Speaker Language Model for Descriptive Profiling**
*Who is this speaker? CoLMbo will tell you.*
[](https://arxiv.org/abs/2506.09375)
[](https://huggingface.co/cmu-mlsp/CoLMbo)
[](https://huggingface.co/datasets/cmu-mlsp/TEARS)
[](https://www.apache.org/licenses/LICENSE-2.0)
[](https://mlsp.cs.cmu.edu)
๐ [Paper](https://arxiv.org/abs/2506.09375) ยท ๐ป [GitHub](https://github.com/massabaali7/CoLMbo) ยท ๐ฆ [TEARS Dataset](https://huggingface.co/datasets/cmu-mlsp/TEARS) ยท ๐ง [Contact](mailto:mbaali@cs.cmu.edu)
---
## What is CoLMbo?
Traditional speaker recognition answers one question: *"Is this speaker A or B?"*
**CoLMbo** asks something richer: *"What is this speaker **like**?"*
Given a few seconds of audio and a natural language prompt, CoLMbo generates free-form descriptions of the speaker their **gender, age, dialect, height, education**, and more directly from voice alone, with no labels or metadata required.
> *"The speaker is a male. He is likely between 26 and 35 years old. He speaks with a New England dialect. He has a Bachelor's Degree."*
CoLMbo integrates a **speaker encoder** with **prompt-conditioned GPT-2 decoding**, enabling zero-shot generalization across diverse speaker populations and datasets.
---
## Table of Contents
- [Quick Start](#quick-start)
- [Example Prompts](#example-prompts)
- [Dataset: TEARS](#dataset-tears)
- [Use Cases](#use-cases)
- [Citation](#citation)
- [Authors](#authors)
---
## Quick Start
### Installation
```bash
pip install transformers torch torchaudio
```
### Load & Run
```python
from transformers import AutoModel
import torchaudio
# Load CoLMbo
model = AutoModel.from_pretrained("cmu-mlsp/CoLMbo", trust_remote_code=True)
model.eval()
# Option A โ from a waveform tensor
waveform, sr = torchaudio.load("speaker.wav")
print(model.describe(waveform, "Please describe the speaker."))
# Option B โ directly from a file path
print(model.describe_file("speaker.wav", "What is the speaker's dialect?"))
```
> โ ๏ธ Audio should be **mono, 16 kHz**. The model will automatically resample if needed.
---
## Example Prompts
CoLMbo accepts any natural language question about the speaker:
```python
prompts = [
"What is the speaker's gender?",
"What is the speaker's age?",
"What is the speaker's dialect?",
"What is the speaker's race?",
"What is the speaker's height?",
"What is the speaker's education level?",
"Please describe the speaker.",
]
for prompt in prompts:
print(f"Q: {prompt}")
print(f"A: {model.describe_file('speaker.wav', prompt)}\n")
```
**Example output:**
```
Q: What is the speaker's gender?
A: The speaker's gender is male.
Q: What is the speaker's age?
A: The speaker is between 26 and 35 years old.
Q: What is the speaker's dialect?
A: The speaker's dialect is from the New England region.
Q: Please describe the speaker.
A: The speaker is a male. He is likely between 26 and 35 years old.
He speaks with a New England dialect. He has a Bachelor's Degree.
```
## Dataset: TEARS
CoLMbo is trained and evaluated on **TEARS** โ a large-scale speaker captioning corpus with rich per-speaker annotations.
๐ฆ **[cmu-mlsp/TEARS on Hugging Face](https://huggingface.co/datasets/cmu-mlsp/TEARS)**
| Split | Utterances |
|:---|:---:|
| Train | 71,100 |
| Test | 44,900 |
| **Total** | **116,000** |
Each example pairs an audio file with a set of `(prompt, response)` pairs covering:
| Attribute | Example Response |
|:---|:---|
| Gender | *"The speaker's gender is male."* |
| Age | *"The speaker is between 26 and 35 years old."* |
| Dialect | *"The speaker's dialect is from the New England region."* |
| Height | *"The speaker is between 5'8 and 5'11."* |
| Education | *"The speaker has a Bachelor's Degree."* |
| Description | *"The speaker is a male. He is likely between 26 and 35..."* |
**Audio sources:**
- [EARS](https://github.com/facebookresearch/ears_dataset) โ expressive, studio-quality single-speaker recordings
- [TIMIT (LDC93S1)](https://catalog.ldc.upenn.edu/LDC93S1) โ phonetically balanced American English speech
---
## Use Cases
| | Use Case | Description |
|:---:|:---|:---|
| ๐ | **Speaker Profiling** | Predict age, gender, dialect, education from voice |
| ๐งฉ | **Explainable Speaker Verification** | Human-readable justifications alongside verification decisions |
| ๐๏ธ | **Speaker-Aware Captioning** | Enrich ASR transcripts with speaker metadata |
| ๐ฌ | **Zero-Shot Attribute Prediction** | Query any speaker attribute without task-specific heads |
| ๐ต๏ธ | **Forensic Audio Analysis** | Generate structured speaker descriptions for investigative use |
---
## Citation
If you find CoLMbo useful in your research, please cite:
```bibtex
@misc{CoLMbo,
title = {CoLMbo: Speaker Language Model for Descriptive Profiling},
author = {Massa Baali and Shuo Han and Syed Abdul Hannan and Purusottam Samal and
Karanveer Singh and Soham Deshmukh and Rita Singh and Bhiksha Raj},
year = {2025},
eprint = {2506.09375},
archivePrefix= {arXiv},
url = {https://arxiv.org/abs/2506.09375},
primaryClass = {cs.CL}
}
```
---
## Authors
Massa Baali ยท Shuo Han ยท Syed Abdul Hannan ยท Purusottam Samal ยท Karanveer Singh ยท Soham Deshmukh ยท Rita Singh ยท Bhiksha Raj
*Carnegie Mellon University โ Language Technologies Institute*
*Machine Learning for Signal Processing Group*
๐ง [mbaali@cs.cmu.edu](mailto:mbaali@cs.cmu.edu)