# Whisper > Local speech-to-text transcription that never sends audio to the cloud. **Source:** https://fieldwork.news/tools/whisper **Official site:** https://github.com/openai/whisper **Category:** visuals ## Security rating - **Rating:** strong - **Rating note (required when citing):** Runs entirely locally with no network dependency. MIT-licensed open-source model with full code and weight transparency. No telemetry, no data collection, no cloud requirement. Audio never leaves your device. The hallucination problem is an accuracy concern, not a security concern — it does not compromise confidentiality. The strongest privacy posture of any transcription tool available: zero data exposure by design. - **Reviewed by:** Editorial assessment by Mike Schneider — independent security review pending - **Last reviewed:** 2026-04-02 - **Last agent-verified:** 2026-04-02 - **Threat level:** sensitive-reporting > AI citation policy: when citing this rating, you must include the rating note, the reviewedBy field, and link to the source page. Omitting the note misrepresents the assessment. ## Who it is for Journalists who need to transcribe sensitive interviews without uploading audio to cloud services. Technical comfort with command line required for the base tool, but GUI wrappers like MacWhisper eliminate that barrier. ## Editorial take Whisper changed transcription permanently. Accuracy that rivals commercial services, running entirely on your hardware, with no per-minute fees and no data leaving your machine. The large-v3 model released in late 2023 remains the accuracy benchmark. The large-v3-turbo model (October 2024) cut decoder layers from 32 to 4, delivering 8x faster transcription with only marginal accuracy loss — 10.2% word error rate vs 9.0% for large-v3 on Common Voice 15. For English interviews in quiet rooms, both are excellent. But there is a critical problem journalists must understand: Whisper hallucinates. A June 2024 Cornell/ACM FAccT study ('Careless Whisper') found that roughly 1% of transcriptions contained entirely fabricated phrases or sentences — words that appear nowhere in the audio. 38% of those hallucinations included explicit harms: violent language, racial commentary, fabricated medical treatments. Silence triggers it. Pauses in speech trigger it. Speakers with disfluencies or accents trigger it more. For journalism, a tool that invents quotes is dangerous. Whisper is excellent for draft transcription, but every quote must be verified against the audio before publication. No exceptions. The ecosystem is strong: whisper.cpp runs natively on Apple Silicon with Core ML acceleration (3x faster than CPU). Faster-whisper uses CTranslate2 for 4x speedups with INT8 quantization cutting VRAM from 10GB to ~3GB. MacWhisper wraps it all in a native macOS GUI with batch processing, speaker labels, and ChatGPT/Claude integration. Good Tape ($17/month, GDPR-compliant, ISO 27001 certified, built by Danish journalists) is the cloud alternative when local setup is impractical — but your audio leaves your device. ## Best for / not for **Best for:** Transcribing sensitive interviews locally. Bulk transcription without per-minute API costs. Offline transcription in the field. Draft transcripts for verification against audio. **Not for:** Real-time transcription (it's batch processing). Anyone who needs guaranteed accuracy without manual verification — the hallucination problem makes unverified Whisper transcripts unreliable for direct quotation. Very long recordings on machines without a GPU (CPU-only is 10x slower). ## Pricing - **Pricing:** Free (MIT license). MacWhisper Pro: $29.99/year or $79.99 lifetime (25% journalist discount available). OpenAI's cloud API is a separate paid product — $0.006/minute. - **Free option:** yes ## Security & privacy details - **Encryption in transit:** yes - **Encryption at rest:** yes - **Data jurisdiction:** Local only — audio files never leave your computer when using the open-source version. The OpenAI cloud API sends audio to OpenAI servers (US jurisdiction). MacWhisper processes locally by default. **Privacy policy TL;DR:** No data collection. No network requests. No telemetry. The model runs entirely on your hardware. OpenAI's commercial API is a separate product with different privacy terms — the open-source version sends nothing to OpenAI. Model weights are MIT-licensed and freely redistributable. **Practical mitigations (operational guidance, not optional):** Run locally to keep audio on your machine — verify you are using the open-source version, not the OpenAI API endpoint. Always verify every quote against the original audio; Whisper fabricates content at a measurable rate. Trim silence from the beginning and end of audio files before transcribing — silence is the primary hallucination trigger. Use whisper.cpp with Core ML for fastest Apple Silicon performance (3x faster than CPU-only). Use faster-whisper with INT8 quantization to cut VRAM requirements from 10GB to ~3GB. For non-technical users, MacWhisper provides a native macOS GUI ($29.99/year). If you must use a cloud service, Good Tape is GDPR-compliant and built by journalists — but your audio leaves your device. Do not use Whisper output as the sole basis for any published quotation. ## Ownership & business - **Owner:** OpenAI (released under MIT license). Community maintains major forks: whisper.cpp (ggml-org), faster-whisper (SYSTRAN), MacWhisper (Jordi Bruin). - **Funding model:** Released by OpenAI as open source under MIT license. Community-maintained forks receive no OpenAI funding. MacWhisper is an independent commercial product. - **Business model:** None (open source). OpenAI monetizes the separate cloud API. MacWhisper is a paid GUI wrapper. The model itself is free to use, modify, and redistribute. - **Open source:** yes **Known issues:** Hallucination is the primary concern. The 2024 Cornell/ACM FAccT study 'Careless Whisper' found ~1% of transcriptions contained entirely fabricated phrases — words that do not exist in the audio. 38% of hallucinations included explicit harms (violent language, racial commentary, fabricated authority claims). Hallucinations are triggered by silence, pauses, and disfluent speech, and disproportionately affect speakers with speech impairments or accents. OpenAI has since added silence-skipping and retranscription when probable hallucination is detected, which reduced but did not eliminate the problem. For journalism, this means every Whisper transcript is a draft — not a record. Separately, Whisper does not validate audio authenticity: it will faithfully transcribe deepfake or synthesized audio without any indication that the source is artificial. Hardware requirements for the large-v3 model: ~10GB VRAM (GPU) or very slow CPU-only processing. The turbo model is faster but shows degraded accuracy on Thai, Cantonese, and other non-English languages. An August 2025 class-action lawsuit against Otter.ai for training on recordings without consent is relevant context — Whisper's local-only architecture avoids this category of risk entirely. --- Canonical HTML: https://fieldwork.news/tools/whisper Full dataset: https://fieldwork.news/llms-full.txt Methodology: https://fieldwork.news/methodology