VALORANT Agent Voice Classifier
Identifies which of the 26 VALORANT agents is speaking in a voice-line recording. It is part of a voice-line QA & localization assistant for Riot Games (HKUST ISOM5240 project): https://github.com/EthanIsaac/valorant-voiceline-qa
- Fine-tuned from:
facebook/wav2vec2-base, selected from 3 candidates on validation accuracy - Data:
NgThVinh/ValorantAgentVoiceLines(revisionc2e0d34), split 80/10/10 per agent, grouped by line - Input: mono audio at 16000 Hz; the first 4 s are used
Results (test set, 1175 clips)
| Metric | Value |
|---|---|
| Accuracy | 0.905 |
| Macro-F1 | 0.905 |
| CPU time per clip | 478 ms |
Usage
from transformers import pipeline
classifier = pipeline("audio-classification", model="eibautistat/valorant-agent-voice-classifier")
classifier("voice_line.wav", top_k=3)
Limitations
Trained on studio-quality in-game voice lines; recordings with music, sound effects or other speakers may be classified less accurately. Agents: astra, breach, brimstone, chamber, clove, cypher, deadlock, fade, gekko, harbor, iso, jett, kayo, killjoy, neon, omen, phoenix, raze, reyna, sage, skye, sova, tejo, viper, vyse, yoru.