Whisper Transcription — Run It Free and Local, or Pay by the Minute

Whisper transcription means converting speech to text with OpenAI's Whisper speech-recognition models. The weights are open source, so you can run the model three ways: through OpenAI's paid API (whisper-1 at \$0.006 per minute), self-hosted in Python, or compiled locally with whisper.cpp - no upload, no per-minute meter, no file cap. Accuracy on clean English audio is roughly 3-6% word error, which is why 'Whisper' has become shorthand for AI transcription quality in general.

Per OpenAI's pricing page, whisper-1 is \$0.006 per minute (\$0.36 per audio hour). The newer menu is cheaper at the floor: gpt-4o-mini-transcribe is \$0.003/min, gpt-4o-transcribe is \$0.006/min with better accuracy, and the realtime streaming variants run \$0.017/min. One catch that matters for subtitle workflows: only whisper-1 still returns word-level timestamps and SRT/VTT output, which is why video pipelines keep using it. At \$0.006/min, 10 hours of monthly audio costs about \$3.60 - metered forever, and every minute uploads to OpenAI's servers.

Yes, permanently, if you run it locally. The open-source weights (MIT license) and the whisper.cpp port are free downloads: a one-time model file (75 MB for the small English model, about 1.5 GB for large-v3-turbo), then unlimited transcription on your own hardware - including on a machine with the ethernet cable pulled. The API route is pay-per-minute. Free web tools that 'use Whisper' typically meter you (TurboScribe 3 files/day, Notta's 3-minute cap) because they run the model on their own GPUs; running it on your PC removes the meter entirely.

Four steps: (1) clone whisper.cpp and build it with Visual Studio's toolchain (cmake -B build, then cmake --build build --config Release), or grab a prebuilt release; (2) download a ggml model - large-v3-turbo is the 2026 accuracy default for English; (3) convert your audio to 16 kHz mono WAV with ffmpeg (ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav); (4) run whisper-cli -m models/ggml-large-v3-turbo.bin -f output.wav -otxt. Add --vad with the Silero model to stop music/silence hallucinations. The whole pipeline runs offline after the model download.

OpenAI's Whisper large-v3 posts 7.44% WER on the public Open ASR Leaderboard; NVIDIA's Parakeet TDT 0.6B v3 posts 6.34% average WER and runs about 45x realtime on CPU - both ahead of Whisper's flagship. large-v3-turbo is the practical default: 809M parameters (the decoder pruned from 32 layers to 4), roughly five times faster than full large-v3 with near-identical accuracy on well-resourced languages. Whisper remains the default choice when you need translation, word timestamps, or SRT export; Parakeet when you need raw English speed.

Depends entirely on which route you pick. The open-source model running locally (whisper.cpp, faster-whisper, or a desktop app) never uploads anything - transcription happens on your CPU/GPU and the only network traffic is the one-time model download. The API route uploads every minute of audio to OpenAI's cloud for processing and billing. Same model, opposite privacy posture. For legal, medical, or client-confidential recordings the local route is the only one where no third party ever receives the audio.

Not vanilla Whisper - it is a batch model. whisper.cpp streams 30-second windows with partial results, which is usable but adds seconds of lag, and OpenAI's whisper-1 endpoint explicitly does not support streaming. OpenAI now sells realtime variants (gpt-realtime-whisper, gpt-live-transcribe at \$0.017/min) for live transcript deltas. Locally, desktop apps built on Whisper-class engines solve this differently: continuous-mode capture transcribes your microphone as a rolling buffer and shows the text with a short delay - good enough for meeting notes and live captions in practice.

VocalFuse runs a Whisper-class engine locally on Windows inside a normal app: dictation into any text field with a global hold-to-talk key, file transcription with timestamps and speaker labels, bot-free meeting capture, and TXT/SRT/VTT export - free tier, Basic \$5/mo, or Pro \$10/mo with AI notes and summaries. Nothing ever uploads because the model runs on your PC after a one-time download. If you want the raw open-source stack instead, whisper.cpp with Silero VAD plus WhisperX for word timestamps and diarization is the free DIY route - both paths end at the same transcript.