Speech-to-Text
Transcribe and understand audio with AI
Speech-to-text models for transcription, meetings, and search
Speech-to-text (STT) models convert spoken audio into written text. The category covers everything from podcast transcripts to real-time captioning pipelines to voice-command interfaces inside mobile apps. Reach for STT when you need to search inside audio, build dictation, summarize meetings, or generate captions for accessibility.
9 models available
Incredibly Fast Whisper
Whisper Large v3 wrapped with Hugging Face Transformers optimizations (batched inference, flash attention) for very high throughput. Transcribes hours of audio in minutes on a single GPU. Maintained by Vaibhav Srivastav. Good when you need bulk transcription fast.
Whisper
OpenAI's Whisper running on Replicate. General-purpose speech recognition trained on 680k hours of multilingual audio. Transcribes and translates 99 languages, robust to accents and background noise, and outputs plain text, segments, or word-level timestamps.
Whisper Large V3
OpenAI's Whisper model. State-of-the-art speech recognition supporting 99+ languages.
Whisper Large v3 Turbo
OpenAI's distilled Whisper Large v3. ~216x realtime, 99+ languages, MIT-licensed weights.
Deepgram Nova-3
Deepgram's flagship STT. First to offer realtime multilingual transcription with self-serve customization.
SeamlessM4T
Meta's SeamlessM4T multimodal translation model. Takes speech or text input and produces transcription or translation across about 100 languages, including speech-to-text and speech-to-speech. One model covers ASR plus cross-lingual translation without chaining separate systems.
SeamlessM4T v2 Large (Speech)
Meta SeamlessM4T v2 Large speech mode. Speech-to-speech, speech-to-text, and text-to-speech translation across 100+ languages in a single unified model.
Whisper Diarization
Whisper Large v3 Turbo combined with pyannote 4.0 for speaker diarization, returning who-said-what segments with timestamps. Built by Thomas Mol. Returns a clean JSON of speaker-labeled segments, handy for meeting notes, interviews, and podcasts.
WhisperX
WhisperX (Large v3) with forced alignment for accurate word-level timestamps plus optional speaker diarization. Uses VAD to cut long files into segments and a wav2vec2 aligner to pin each word to its exact time. Useful for subtitles and per-speaker transcripts.
Top speech-to-text picks
Hand-picked across four common criteria — resolved against the live catalog so the picks track price and performance changes.
Whisper Large v3 wrapped with Hugging Face Transformers optimizations (batched inference, flash attention) for very high throughput. Transcribes hours of audio in minutes on a single GPU. Maintained by Vaibhav Srivastav. Good when you need bulk transcription fast.
Learn moreDeepgram's flagship STT. First to offer realtime multilingual transcription with self-serve customization.
Learn moreWhisper Large v3 wrapped with Hugging Face Transformers optimizations (batched inference, flash attention) for very high throughput. Transcribes hours of audio in minutes on a single GPU. Maintained by Vaibhav Srivastav. Good when you need bulk transcription fast.
Learn moreOpenAI's Whisper model. State-of-the-art speech recognition supporting 99+ languages.
Learn morePricing is almost always per-minute of audio. Flagship models (Whisper Large V3, Deepgram Nova-3, ElevenLabs Scribe) cost roughly €0.005-€0.015 per minute. A one-hour podcast transcript costs €0.30-€0.90 depending on the tier. Some providers charge extra for premium features like speaker diarization, word-level timestamps, summaries, or translation, so do the math with the features you actually need turned on.
The trade-off is accuracy, latency, and feature richness. Whisper Large V3 leads on raw word-error-rate in benchmark evaluations and is open-weights, so you can self-host. Deepgram Nova-3 and AssemblyAI Universal lead on streaming latency (sub-300ms first token) and diarization quality. ElevenLabs Scribe leads on multilingual coverage and code-switching (when speakers swap languages mid-sentence). For batch transcription, Whisper usually wins on cost-and-accuracy. For realtime call transcription, a streaming-first provider wins.
Watch out for noisy audio: word-error-rate roughly doubles below 20 dB SNR on every model, and overlapping speakers degrade diarization even on the flagships. Pre-process with a noise-suppression model (RNNoise, Krisp) if your source is unpredictable. Also watch out for proper nouns: every model still mistranscribes uncommon names, technical terms, and brand names. Most providers accept a `keywords` hint list to bias the decoder — use it.
Top picks above cover the most accurate model, the cheapest workhorse, the longest-audio supporter, and the fastest streaming option.
Popular use cases
Common patterns built with speech-to-text on Railwail.
Related comparisons
Side-by-side reviews of the most-compared models in this category.
Frequently asked questions
Start Building with AI
Access all models through a single API. Get free credits when you sign up — no credit card required.