Auto Captions & Whisper-based Subtitles

The "drop a video in, get an SRT out" workflow. As of 2026 a Whisper-class model running locally on a laptop is good enough for most YouTube / podcast / corporate caption needs. For full STT systems and hosted APIs see Speech & Transcription. For hand-editing the result see Subtitles. For NLE picks see Video NLE Editors.

Local Whisper engines

• ★ whisper.cpp — MIT OSS C/C++ port; runs on CPU, Metal (Apple Silicon), CUDA, Vulkan, OpenCL. The fastest practical local Whisper in 2026 — large-v3 on an M-series Mac runs faster than realtime. Default for command-line and embedded use.
• ★ faster-whisper — MIT OSS Python; CTranslate2-backed; ~4× faster than vanilla openai-whisper with the same accuracy. Default for Python pipelines and server-side batch.
• ★ WhisperX — BSD OSS Python; faster-whisper + word-level timestamps via wav2vec2 forced alignment + speaker diarization (pyannote). The right choice when you need accurate per-word timing for kinetic captions.
• openai-whisper — MIT OSS reference Python; slower than the alternatives but reference accuracy.
• insanely-fast-whisper — Apache OSS; HF Transformers + Flash Attention; very fast on a CUDA GPU.
• distil-whisper — MIT OSS distilled English-only Whisper; ~6× faster, slight accuracy hit; great for English podcasts.

Desktop GUIs (point-and-click)

• ★ MacWhisper — closed source, paid (~$60 lifetime), free tier limited. The polished Mac front-end for whisper.cpp. Drag a file in, get SRT/VTT/TXT/MD out. Default Mac pick if you don't want CLI.
• ★ Buzz — MIT OSS; cross-platform Qt app; whisper.cpp + faster-whisper backends; free; the OSS counterpart to MacWhisper.
• WhisperWriter — GPL OSS; Windows-focused live dictation.
• Subtitle Edit — see Subtitles; has whisper.cpp / faster-whisper / Purfview's Whisper baked into "Audio to text" menu.
• Aiko (macOS) — closed source, free; very simple drag-and-drop Whisper app.
• Whisper Transcription (Mac App Store) — paid, polished.

NLE / video-editor integrations

• DaVinci Resolve Studio — paid; built-in Speech-to-text (uses internal model, not Whisper, but comparable).
• Premiere Pro — bundled Adobe Sensei caption generation; subscription.
• Kdenlive — GPL OSS; whisper.cpp integrated since 23.04; right-click clip → Speech to Text.
• Descript — paid free tier; integrated transcription drives the editor itself.

Mobile / social-video AI captions

• ★ CapCut Mobile auto-captions — free; ByteDance — flag: data leaves your device. Useful for personal social posts; avoid for confidential.
• Submagic — paid; AI captions + emoji + b-roll for TikTok / Reels / Shorts.
• Captions (the app) — paid; AI talking-head editor with caption styling.
• Veed — paid free tier; web-based auto-captioning.
• Riverside Magic Clips — paid; same.

Hosted Whisper / speech APIs

• See Speech & Transcription for Deepgram, AssemblyAI, OpenAI Whisper API, Speechmatics, Gladia, Soniox.
• OpenAI Whisper API — cheap; no diarization; no word timestamps unless you ask.
• Replicate — pay-per-second hosted Whisper / WhisperX models.
• AssemblyAI — paid free tier; best accuracy + diarization in 2026 for hosted.
• Deepgram Nova — paid free credits; not Whisper-based but competitive.

Live / streaming captioning

• ★ faster-whisper-server / WhisperLive — OSS WebSocket servers that stream Whisper transcriptions live; pair with OBS browser source for on-stream captions.
• OBS + Web Captioner — free closed; uses browser STT; OBS browser source overlays the captions.
• Cloud Captioner — paid SaaS for live events.
• StreamerSonglist + caption bot — Twitch-side patterns.
• Stenofy / 1CapApp — paid live human / hybrid captioning.

Model size guidance (2026)

• tiny / base — phone / Pi-class; ok for clean English; struggles on accents.
• small — laptop CPU realtime; fine for podcasts.
• medium — sweet spot for accuracy/speed on M-series / a 4060.
• ★ large-v3 — best general accuracy; runs realtime on Apple Silicon and recent NVIDIA.
• large-v3-turbo — 2024 OpenAI release; ~8× faster than v3 with minor accuracy loss; default on whisper.cpp now.
• Language-tuned forks — NB Whisper (Norwegian), Whisper-Fa (Persian), Belle-Whisper (Chinese) — better than vanilla on those languages.

Pipeline tips

• Convert to 16 kHz mono WAV first with ffmpeg -i in.mp4 -ar 16000 -ac 1 out.wav for fastest inference.
• Run faster-whisper with vad_filter=True to drop silences — faster + fewer hallucinated lines.
• Use WhisperX when you need word-level timing for kinetic / karaoke-style captions.
• Burn in vs soft subs — keep soft (.srt / .vtt sidecar) for web / accessibility; burn-in only for fixed social formats.
• Spot-check the result in Subtitle Edit — Whisper hallucinates on silence and music; trim those lines.

Pick this if…

• Default local CLI: whisper.cpp.
• Default Python pipeline: faster-whisper.
• Need word-level timing or speaker labels: WhisperX.
• Mac, GUI, ok with paid: MacWhisper.
• Mac/Win/Linux GUI, free: Buzz.
• Already in Subtitle Edit: built-in "Audio to text".
• Live on-stream captions: WhisperLive + OBS browser source.
• Best hosted accuracy: AssemblyAI or Deepgram Nova.
• Mobile social, casual: Submagic or CapCut (with the ByteDance flag).