Whisper and ASR Models

Whisper and ASR models

Whisper is OpenAI’s open-weight ASR model; per OpenAI’s docs (2026-10-02) its hosted transcription recommends newer gpt-* models.

Whisper (open source)

• openai’s repo: code and weights under the MIT licence (about 110k GitHub stars as of 2026-10-05; last tagged release v20250625, GitHub API 2026-10-05). Sizes: tiny 39M, base 74M, small 244M, medium 769M, large 1550M, turbo 809M (an optimised large-v3: faster, small accuracy loss, cannot translate).
• Supports multilingual recognition, translation to English and language identification. No Whisper generation beyond large-v3 / large-v3-turbo exists under the openai organisation on Hugging Face (newest checkpoint large-v3-turbo, 2024-10-04; HF API 2026-10-05).

OpenAI hosted transcription (docs, 2026-10-02)

• gpt-transcribe: recommended general transcription model; supports prompt, keyword and language hints.
• gpt-4o-transcribe-diarize: speaker diarization with timestamps and known-speaker references.
• whisper-1: “legacy”; still needed for word/segment timestamps and translation (gpt-transcribe does not translate yet). gpt-4o-transcribe and gpt-4o-mini-transcribe remain available.
• Models overview also lists GPT-Realtime-Whisper and the gpt-live / gpt-realtime voice models. File limit 25 MB.

Other current ASR

• NVIDIA Parakeet-TDT-0.6B-v3 (Nvidia): 600M parameters, 25 European languages, CC-BY-4.0, average WER 6.34% on the Open ASR leaderboard as stated by the model card (not re-verified); model repo created 2025-08-04 on Hugging Face (licence CC-BY-4.0 confirmed).
• voxtral (Mistral) and lfm2 (audio-native models that also do ASR).
• Leaderboard: Hugging Face Open ASR Leaderboard (contents did not load this session; rankings unverified).

Choosing

Suggestions (opinion): local/private use Whisper turbo or Parakeet; hosted with diarization gpt-4o-transcribe-diarize; consider post-processing domain terms with a text model.

Open items

• Open ASR Leaderboard rankings and Parakeet WER (6.34%) not re-verified; Voxtral/LFM2 claims are only link references.

Sources (accessed 2026-10-02)

• https://github.com/openai/whisper
• https://developers.openai.com/api/docs/guides/speech-to-text
• https://huggingface.co/api/models?author=openai&search=whisper (accessed 2026-10-05)
• https://developers.openai.com/api/docs/guides/speech-to-text (accessed 2026-09-30, not re-fetched)
• https://developers.openai.com/api/docs/models (accessed 2026-09-30, not re-fetched)
• https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 (accessed 2026-09-30, not re-fetched)