Speech-to-Text API

A speech-to-text API turns audio into text. The four models that matter on Replicate cover four distinct needs: the reference transcriber, the fast one, the most accurate one, and the one that knows who said what.

OpenAI Whisper (144M+ runs) is the reference. Incredibly Fast Whisper (38M+ runs) trades features for raw speed. GPT-4o Transcribe brings OpenAI's newer architecture with prompt-biased vocabulary. Whisper Diarization (8.6M+ runs) adds speaker labels.

Quick Recommendation

For most workloads, start with vaibhavs10/incredibly-fast-whisper — in example runs it transcribes in ~3s what standard Whisper takes ~17s to process, with timestamps included.

Use openai/whisper when you need the full feature set: translation to English, temperature control, and the most battle-tested behavior on Replicate (144M+ runs).

Use openai/gpt-4o-transcribe when accuracy on domain vocabulary matters — its prompt input biases the model toward names, jargon, and acronyms it would otherwise mis-hear. Note it returns clean text without word-level timestamps.

Use thomasmol/whisper-diarization when you need who said what — meetings, interviews, podcasts — at the cost of longer runs (~52s in the cached example).

Keyword-Matched Use Cases

Model Comparison

Choosing in practice

Speed: Incredibly Fast Whisper batches audio through an optimized pipeline — for long files the gap versus standard Whisper is dramatic. If transcription latency sits in a user-facing flow, this is the default.

Accuracy on hard vocabulary: all Whisper variants share training-data blind spots on niche names and jargon. GPT-4o Transcribe's prompt input is the only lever here — feed it product names, speaker names, or domain terms and error rates on those words drop.

Speaker separation: none of the plain transcribers label speakers. Whisper Diarization combines transcription with diarization in one call — slower, but far simpler than stitching two models yourself.

Translation: classic Whisper transcribes ~100 languages and can translate to English in the same run (translate flag). GPT-4o Transcribe takes a language hint instead.

Practical Decision

Default to Incredibly Fast Whisper for speed and cost. Escalate to GPT-4o Transcribe when domain vocabulary or punctuation quality matters more than timestamps. Reach for Whisper Diarization the moment "who said what" appears in the requirements, and keep classic Whisper for translation workflows.

FAQ

What is the best speech-to-text API on Replicate?

Incredibly Fast Whisper for most jobs — Whisper-level output at a fraction of the runtime. Classic Whisper remains the most used (144M+ runs) and adds translation.

Which transcription API gives speaker labels?

thomasmol/whisper-diarization — it transcribes and labels speakers in one call. Plain Whisper and GPT-4o Transcribe do not identify speakers.

Does GPT-4o Transcribe give timestamps?

No — it returns clean punctuated text. For subtitles or alignment, use a Whisper variant with timestamps.

How do I improve accuracy on names and jargon?

Use GPT-4o Transcribe and pass the relevant terms in its prompt input. That biases recognition toward your vocabulary — something classic Whisper cannot do.