Observed Signal · Jun 25, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
ASR Hallucinations: Phantom 'Thank you' in Meeting Transcripts
A developer of the passive AI meeting assistant Faktum investigated repeated, spurious "Thank you." lines appearing in transcripts. The phantom phrases occurred not from mis-heard speech but whenever the audio contained applause, laughter, or silent gaps — a learned association in ASR models trained on real-world audio. The streaming transcription API provided no confidence or no-speech probability to filter by, and energy-based VAD tuning could not distinguish applause from speech. The author implemented a conservative post-processing exact-match blocklist (TranscriptHallucinationFilter) that drops known multi-word hallucination phrases at the ingestion chokepoint, logs drops for monitoring, and acknowledges trade-offs (English-first, occasional dropped genuine isolated thanks).
Practical engineering fix improves transcript quality for AI meeting assistants and similar ASR-driven products; important for product reliability but not industry-shifting.
Track Real-Time Speech Recognition / ASR Signals & Market Shifts
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Author builds Faktum, a passive AI meeting assistant that records, transcribes live in 26 languages, and fact-checks claims in real time.
- Phantom transcript lines (e.g., repeated "Thank you.") aligned with applause, laughter, or silence — not actual speech.
- ASR models (e.g., Whisper, qwen3-asr) learn to emit common human transcriptions paired with non-speech audio (clapping → "Thank you.").
- The streaming transcription events lack confidence scores or a no_speech_probability field, preventing confidence-threshold filtering.
- The implemented fix is a conservative post-processing exact-match blocklist (TranscriptHallucinationFilter) applied before saving transcripts and feeding downstream systems.
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Catch AI Hallucinations Before Your Audience Does
The author describes a repeatable validation system for catching AI-generated content hallucinations before publication. After reviewing roughly 400 AI-assisted pieces, they identify three common hallucination patterns—laundered statistics, misattributed quotes, and stale facts presented as current—and recommend a lightweight verification stack (Perplexity AI, Google Scholar, QuoteInvestigator, news search) plus a prompt-based technique (
System Prompt Reduces AI Hallucinations
t3n published guidance and a reusable system-prompt template (via its t3n MeisterPrompter podcast and newsletter) aimed at reducing hallucinations from AI chat tools. The prompt instructs models to explicitly declare uncertainty (e.g., say “I don't know”), avoid inventing facts, sources or numbers, and label assumptions. The article explains where to set system prompts in common assistants (ChatGPT, Claude, Google Gemini) and notes that while a system prompt helps detect and reduce errors, it cannot fully prevent hallucinations. The full prompt text is available in the podcast show notes and related newsletter materials.
How to Diagnose and Reduce AI Coding Agent Hallucinations
A Dev.to technical post (published 2026-05-23) explains why AI coding agents hallucinate and offers a practical feedback loop to reduce repeated errors. The author advises engineers to diagnose what the agent wrongly invented, trace the context sources that influenced the decision (conversation history, repo-level rules like CLAUDE.md/AGENTS.md, and automatic memory), and then fix those inputs rather than only correcting outputs. Recommended tactics include context isolation (moving niche rules into Skills/Subagents), pruning or editing automatic memories, and treating agent context as living code that requires refactoring and testing. The piece cites research showing models are rewarded to guess rather than admit uncertainty and emphasizes that hallucinations cannot be eliminated but can be reduced and recovered from faster.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.