VOICE 2021 | Daria Baidakova, Toloka
summary
Crowdsourcing scales collection and still requires humans for training, evaluation and client benchmarks; recordings under instruction use local volunteers, trained labelers and automated ASR prompts that speed work but introduce bias. Tasks split into short clips — keep clips <5s (10s already strains quality) — with dynamic overlap stopping after X exact matches, labeler qualification and aggregation. Quality is measured by word error rate (WER), recall, accuracy, mean overlap, METEOR and BLEU and by per-clip cost.
AI-written summary · about
chapters (3)
related talks
21mA Human-in-the-Loop System for Sound Event Detection and Annotation
17mYour AI Deserves Better Data
17mBuilding Effective Voice Agents
20mFrom Self-driving to Autonomous Voice Agents
16mServing Voice AI at $1/hr: Open-source, LoRAs, Latency, Load Balancing
43mGiving my wife a voice with Ruby and AI cloning (+ discussion)
39mVideo Generation + AI: From Production to Product
01h 09mAI is eating the world (and Google)