Whisper

  • Whisper (OpenAI Whisper) — Open-Source Multilingual ASR Baseline + $0.006/min Hosted API

    OpenAI Whisper launched Sep 2022 (MIT-licensed weights on GitHub), and as of 2026-08 the open-source repo is intact, whisper-1 is still callable on the OpenAI API at $0.006/min, and OpenAI has added GPT-4o Transcribe / Mini Transcribe / Diarize as newer hosted siblings on the same /v1/audio/transcriptions path — Whisper itself was not renamed, merged, or retired. Category ai-productivity-data; pipeline_stage none. It is a model + reference runtime, not a consumer app: no dashboard, no login, no subscription from OpenAI. Killer differentiator = only mainstream ASR you can run fully offline under MIT with 99-language coverage; not Deepgram (real-time-first), not AssemblyAI (managed intelligence layer), not Otter/Descript (end-user products), not gpt-4o-transcribe (higher accuracy, no SRT/VTT). 2026 trap: base Whisper hallucinates on silence, no native diarization/real-time, 25MB/25-min API cap, large-v3 needs ~10GB VRAM.

    1. One-line positioning

    Whisper = Transformer ASR trained on 680k hours multilingual audio, chunked 30-sec, outputs text/json/srt/vtt with segment+word timestamps; runs as openai/whisper (PyTorch, MIT) or via OpenAI API whisper-1 at $0.006/min, or as community runtimes whisper.cpp / faster-whisper / WhisperX. Fits ai-productivity-data; pipeline_stage none (audio→chunk→logprob decode→transcript+ts). For batch transcription where you want zero per-minute cost + full data control, it’s the default baseline; for live captioning or speaker-labeled meetings, it’s the wrong primitive without wrappers.

    2. Core features

    Layer Capability Notes
    Model family tiny(39M)/base(74M)/small(244M)/medium(769M)/large-v3(1550M)/large-v3-turbo(809M, ~8× speed)​ FP32 reference; FP16/INT8 via ports
    Languages 99 languages​ + translate-to-English; strong on accents/noise/code-switching (trained on web audio, not studio) Quality uneven on low-resource langs
    Outputs (self-host) text / json / verbose_json / srt / vtt / tsv; segment + optional word timestamps (whisper_timestamped)
    Outputs (API whisper-1) json/text/srt/verbose_json/vtt + timestamp_granularities[] (word/segment) Only Whisper-1 keeps SRT/VTT; gpt-4o-transcribe returns json/text only
    Diarization None native​ — bolt on pyannote.audio, or use WhisperX (faster-whisper + pyannote + wav2vec2 forced align)
    Real-time Batch-only in reference impl; whisper.cpp / faster-whisper can stream with wrappers; API whisper-1 is not real-time​ (use gpt-realtime-whisper $0.017/min)
    Runtimes openai/whisper (PyTorch) · whisper.cpp (C++/ggml, CPU+Metal, edge) · faster-whisper (CTranslate2, 2–4× faster) · WhisperX (diarization+alignment) · insanely-fast-whisper (FlashAttn2)
    Hosted API whisper-1 $0.006/min · gpt-4o-mini-transcribe $0.003/min · gpt-4o-transcribe $0.006/min · gpt-4o-transcribe-diarize $0.006/min (diarized_json) 25MB/file, 1500s/req cap

     

    3. Pricing (2026-08)

    Path Price Audience
    Self-hosted (MIT)​ $0​ + your GPU/CPU compute Privacy-first, batch volume, offline/edge
    OpenAI whisper-1 API $0.006/min​ ($0.36/audio-hr) Existing Whisper integrations, SRT/VTT need
    OpenAI gpt-4o-mini-transcribe $0.003/min ($0.18/hr) Cheapest managed batch, json/text only
    OpenAI gpt-4o-transcribe $0.006/min ($0.36/hr) Better accuracy, json/text only
    OpenAI gpt-4o-transcribe-diarize $0.006/min Speaker labels (non-realtime HTTP)
    OpenAI gpt-realtime-whisper $0.017/min ($1.02/hr) Live captioning / voice agents
    Third-party hosted (Groq/Replicate/Fireworks/Together) $0.02–0.10/hr (Whisper weights) High-throughput batch, no OpenAI lock

    pricing_type = open-source-free-or-api-per-minute; starting_price = $0 (self-host); free_tier = "yes; MIT weights self-hosted free, no OpenAI free tier on API (pay from first min)"; pricing_note = "whisper-1-0.006-min-36-hr-25mb-1500s-cap; gpt-4o-mini-transcribe-0.003-min-json-text-only; gpt-4o-transcribe-0.006-min-json-text-only; diarize-same-price-diarized-json; gpt-realtime-whisper-0.017-min; large-v3-turbo-809M-8x-speed; no-native-diarization-no-native-realtime-in-reference; groq-whisper-turbo-0.02-hr-alt"

    4. Access Type

    access_type: open-source-model-plus-api

    access_display: 💻 openai/whisper (Python/CLI) · 🌐 OpenAI /v1/audio/transcriptions · 🔌 whisper.cpp/faster-whisper/WhisperX runtimes · (no official OpenAI consumer UI)

    5. Reviews (2026)

    Source Signal
    GitHub (openai/whisper) ~75k stars; reference baseline everyone forks; complaints = slow ref impl, hallucination on silence, no diarization
    G2 / tool roundups Not a SaaS, so no G2 entry; in dev surveys ~4.4/5 as model, dinged as product
    AISO / AI Pedia / Dev.to 2026 “Best offline multilingual baseline, at $0 if you have a GPU; but production STT usually wraps it or picks Deepgram/AssemblyAI for realtime+diarization”
    Trustpilot / consumer N/A — users review Otter/Descript, not Whisper itself

     

    Praise: MIT weights forever free, 99-lang coverage unmatched in open source, robust on noisy/accented speech, huge ecosystem (whisper.cpp/faster-whisper/WhisperX/distil-whisper), SRT/VTT only on whisper-1 among OpenAI STT models.

    Gripes: hallucinates repeated phrases on dead air, no native speaker labels, reference impl slow on CPU, large-v3 needs ~10GB VRAM, 30-sec chunking hurts long-form coherence, API file/duration caps force client-side splitting, no built-in punctuation tuning/summarization/PII redaction.

    6. Best for / Not for

    Best for Not for
    Batch transcription of podcasts/lectures/meetings/interviews (post-hoc) Live captioning / voice agents needing <300ms first token (use Deepgram/gpt-realtime-whisper)
    Privacy-bound orgs (legal/medical/research) wanting audio to never hit a cloud Multi-speaker meetings needing “who said what” out of the box (needs WhisperX/pyannote)
    Developers building custom STT pipelines, already on OpenAI stack Non-technical users wanting a record-button app (use Otter/Descript/TurboScribe)
    Low-resource / non-English / code-switching audio High-volume call-center telephony (Deepgram Nova-3 WER better on phone audio)
    Edge/on-device (whisper.cpp on Apple Silicon/RPi) Teams wanting summarization/sentiment/PII bundled (AssemblyAI)

     

    7. Competitors

    Tool Lane Starts
    Deepgram Nova-3 Real-time STT, native diarization, ~300ms $0.26/min? batch $0.0036/min, stream $0.0056/min
    AssemblyAI Universal-2/3 Managed STT + summarization/sentiment/PII $0.12–0.75/hr, 5h free
    gpt-4o-transcribe / mini OpenAI’s higher-accuracy hosted Whisper successor (json/text only) $0.003–0.006/min
    Groq Whisper large-v3-turbo Hosted Whisper, LPU speed ~$0.02/hr
    Otter.ai / Descript / TurboScribe End-user transcription apps Freemium / $10–20
    Google Chirp 3 / AWS Transcribe / Azure Speech Cloud STT suites $0.0048–0.024/min
    whisper.cpp / faster-whisper / WhisperX Open-source runtimes (not competitors, ecosystem) $0

     

Do Not Sell or Share My Personal Information Cookie Settings