Model Card for voxtral-sentinel-4b

This model is a fine-tuned version of mistralai/Voxtral-Mini-4B-Realtime-2602. It has been trained using TRL.

trishtan/voxtral-sentinel-4b

voxtral-sentinel-4b is a fine-tuned version of mistralai/Voxtral-Mini-4B-Realtime-2602, specialised for real-time audio understanding in high-stakes operational environments. Given a raw audio recording, the model produces a structured output containing a verbatim transcript, a contextual analysis of speaker emotion and situation, and a recommended action — enabling autonomous routing and triage without human-in-the-loop intervention.

Built for two primary verticals:

• Automated customer support — classify caller intent and emotional state to route calls, trigger escalations, or generate automated responses in real time
• Emergency services & safety — identify distress, urgency, and situational context from audio to assist dispatchers or fully autonomous response systems

---

Model Details

Property	Value
Base model	mistralai/Voxtral-Mini-4B-Realtime-2602
Model type	Audio-to-text (multimodal)
Parameters	~4B
Fine-tune method	Full fine-tune (no LoRA)
Precision	bfloat16
Training hardware	NVIDIA A100
Framework	Transformers + TRL SFTTrainer
Language	English
License	See base model license

---

Training

Dataset

Fine-tuned on a curated dataset of ~9,984 audio samples with structured annotations voxtral-forensic-ds. Each sample consists of a raw audio clip paired with a ground-truth output in the following canonical format:

### TRANSCRIPT:
<verbatim transcription of the audio>

### ANALYSIS:
<contextual analysis of speaker emotion, tone, and situation>

### CONCLUSION:
<recommended action or classification>

The dataset was derived from MELD (Multimodal EmotionLines Dataset), which contains emotionally rich conversational audio from multi-speaker dialogue scenarios, and DCASE 2025 Task 1 (Acoustic Scene Classification). Annotations were generated and standardised using automated pipelines with LLM-assisted formatting normalisation.

A 90/10 train/eval split was used with a fixed seed (42) for reproducibility. The final training dataset and held-out eval split are available at trishtan/voxtral-forensic-ds-splits.

Hyperparameters

Parameter	Value
Epochs	5 (early stopping at eval loss < 1.15)
Learning rate	5e-6
LR scheduler	Cosine
Warmup ratio	0.05
Batch size (per device)	2
Gradient accumulation steps	4
Effective batch size	8
Max grad norm	1.0
Precision	bf16
Eval strategy	Every 100 steps

Training Results

Metric	Value
Final eval loss	1.148
Final eval mean token accuracy	74.35%
Train/eval accuracy gap	~0%
Stopped at epoch	2.75 (early stopping)

The near-zero gap between train and eval accuracy across all runs indicates the model generalises well to unseen audio with no measurable overfitting.

---

Usage

import torch
import soundfile as sf
import numpy as np
from transformers import AutoProcessor, VoxtralRealtimeForConditionalGeneration

model_id = "trishtan/voxtral-sentinel-4b"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = VoxtralRealtimeForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Load your audio (must be 16kHz mono)
audio, sr = sf.read("your_audio.wav")
if audio.ndim > 1:
    audio = audio.mean(axis=1)
audio = audio.astype(np.float32)

PROMPT = "[INST] Analyze this recording for forensic indicators. [/INST]"

audio_inputs = processor.feature_extractor(
    [audio], sampling_rate=16000, return_tensors="pt", padding=True,
)
text_inputs = processor.tokenizer(
    [PROMPT], return_tensors="pt", padding=True,
)
inputs = {**audio_inputs, **text_inputs}
inputs = {k: v.to(model.device) for k, v in inputs.items()}

with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)

response = processor.tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(response)

Expected Output Format

### TRANSCRIPT:
I need help immediately, my neighbour hasn't responded in hours and I can hear something...

### ANALYSIS:
The speaker exhibits elevated vocal stress indicators including increased speech rate and
pitch variance. Tone suggests genuine distress rather than rehearsed or non-urgent
communication. Situational context implies potential welfare concern for a third party.

### CONCLUSION:
Escalate to emergency services. Flag as high-priority welfare check. Do not route to
standard support queue.

---

Intended Use

In Scope

• Real-time audio triage in customer service pipelines
• Emergency call classification and dispatcher assistance
• Automated sentiment and intent detection from voice
• Proof-of-concept and research into multimodal audio understanding

Out of Scope

• Medical diagnosis or clinical decision-making
• Surveillance or non-consensual audio analysis
• Languages other than English
• Audio clips under 3 seconds (insufficient signal for reliable analysis)

---

Limitations

• Short audio clips — clips under 3 seconds are padded with silence to the model's required 15-second input window. Analysis quality degrades significantly for very short recordings.
• Single-language — trained exclusively on English-language audio. Performance on accented, non-native, or non-English speech is untested.
• Emotional diversity — training data skews toward conversational emotional registers. Performance on domain-specific audio (medical, legal, industrial) may vary.
• Not a safety-critical system — outputs should be reviewed by human operators in any deployment where errors have real-world consequences.

---

Data Attribution

This model was fine-tuned using audio data derived from:

MELD — Multimodal EmotionLines Dataset Poria, S., Hazarika, D., Majumder, N., Naik, G., Cambria, E., & Mihalcea, R. (2019, July). Meld: A multimodal multi-party dataset for emotion recognition in conversations. In Proceedings of the 57th annual meeting of the association for computational linguistics (pp. 527-536). HuggingFace: ajyy/MELD_audio

DCASE 2025 Challenge — Task 1: Acoustic Scene Classification Annamaria Mesaros, Toni Heittola, and Tuomas Virtanen. A multi-device dataset for urban acoustic scene classification. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2018 Workshop (DCASE2018), 9–13. November 2018. URL: https://dcase.community/documents/workshop2018/proceedings/DCASE2018Workshop_Mesaros_8.pdf.

---

Framework Versions

• TRL: 0.29.0
• Transformers: 5.2.0
• PyTorch: 2.10.0
• Datasets: 4.6.1
• Tokenizers: 0.22.2

---

Citation

@misc{voxtral-sentinel-4b,
  author    = {trishtan},
  title     = {voxtral-sentinel-4b: Fine-tuned Voxtral for Audio Triage},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/trishtan/voxtral-sentinel-4b}
}

@inproceedings{poria2019meld,
  title={Meld: A multimodal multi-party dataset for emotion recognition in conversations},
  author={Poria, Soujanya and Hazarika, Devamanyu and Majumder, Navonil and Naik, Gautam and Cambria, Erik and Mihalcea, Rada},
  booktitle={Proceedings of the 57th annual meeting of the association for computational linguistics},
  pages={527--536},
  year={2019}
}

@inproceedings{Mesaros2018_DCASE,
    Author = "Mesaros, Annamaria and Heittola, Toni and Virtanen, Tuomas",
    title = "A multi-device dataset for urban acoustic scene classification",
    year = "2018",
    booktitle = "Proceedings of the Detection and Classification of Acoustic Scenes and Events 2018 Workshop (DCASE2018)",
    month = "November",
    pages = "9--13",
    keywords = "Acoustic scene classification, DCASE challenge, public datasets, multi-device data",
    url = "https://dcase.community/documents/workshop2018/proceedings/DCASE2018Workshop\_Mesaros\_8.pdf"
}

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}

---

Acknowledgements

Built on Voxtral-Mini-4B-Realtime by Mistral AI. Fine-tuning infrastructure: HuggingFace Transformers, TRL, and Accelerate.