How to choose a speech-to-text model
Transcription models are judged on accuracy (word error rate), languages, speed and price per minute of audio. This guide walks through each, with live prices from our catalogue.
Speech-to-text models by the numbers
Quick answer: top picks today
The three highest-rated models you can call on LLM API right now.
Find your model in 3 questions
1. What matters most?
2. Where will it run?
3. Budget?
Your top 3 update as you answer, from live catalogue data.
Price vs rating
Each dot is a model you can use. Up is better rated, left is cheaper. Hover for details, click to open.
Highest-rated live models
What will it cost per month?
Drag the slider to your expected volume. Prices are live list prices of the top models.
| Model | Per month |
|---|
Key numbers to know
| 1 minute of speech | ≈ 150 spoken words |
| 1 hour of audio as text | ≈ 9,000 words, ≈ 12K tokens |
| Word error rate (WER) 5% | ≈ 1 wrong word per sentence of 20 |
| Typical price range | $0.001–$0.01 per minute |
| Real-time factor | Batch models often transcribe 1 hour in under a minute |
How to choose, step by step
Batch or live?
Recorded files (calls, podcasts, meetings) can use any model. Live captions and voice agents need streaming and a first result in under 500 ms.
Check word error rate on audio like yours
Lower WER is better. Clean podcast audio can hit 3–5%; noisy phone calls often land at 10–20%. Test with 10 of your own recordings before deciding.
Languages and accents
Whisper-family models cover 90+ languages. Some newer models are sharper in English but weaker elsewhere.
Extras you may need
Timestamps, speaker labels (diarisation) and punctuation vary by model. Missing features cost engineering time later.
Price per minute vs self-hosting
Open models like Whisper large-v3 are free to run but need a GPU. Above roughly 50,000 minutes a month, self-hosting often wins.
Decision diagram
Answer top to bottom. The first yes is your starting point.
What the research says
The independent benchmarks that matter for this category, with links to the papers and methodology behind each one.
Hugging Face Open ASR Leaderboard ↗
Average word error rate across 8 English datasets plus speed
Paper: Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation (2025) ↗Artificial Analysis Speech-to-Text ↗
Independent WER, speed and price per 1,000 minutes for hosted APIs
Methodology: Artificial Analysis Speech-to-Text benchmarking ↗Common Voice & FLEURS ↗
Public multilingual test sets used to compare accuracy across languages
Paper: FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech (Conneau et al., 2022) ↗Reviews: what the labs and sources say
Verbatim quotes from public sources. Click through to read them in context.
“In FLEURS & Common Voice benchmark tests across 99 languages, it consistently outperforms leading models like Gemini 2.0 Flash, Whisper Large V3 and Deepgram Nova-3.”ElevenLabs Scribe v1 · ElevenLabs, Meet Scribe, Feb 26, 2025
“Nova-3 delivers industry-leading performance with a 54.3% reduction in word error rate (WER) for streaming and 47.4% for batch processing compared to competitors.”Deepgram Nova 3 · Deepgram, Introducing Nova-3, Feb 12, 2025
“We’re introducing new gpt-4o-transcribe and gpt-4o-mini-transcribe models with improvements to word error rate and better language recognition and accuracy, compared to the original Whisper models.”GPT-4o Mini Transcribe · OpenAI, Introducing next-generation audio models in the API, Mar 20, 2025
“It offers improvements to word error rate and better language recognition and accuracy compared to original Whisper models.”GPT-4o Transcribe · OpenAI API documentation, 2025
Mistakes to avoid
Trusting a single WER number. Accuracy on clean audiobooks says little about noisy phone calls.
Forgetting diarisation. Adding speaker labels later often means switching models.
Paying per minute for silence. Trim long silences before sending audio.
Using a batch model for a voice agent. Latency matters more than the last 1% of accuracy there.
FAQ
What is word error rate (WER)?
The share of words the model gets wrong (substituted, missed or added). 5% WER means about 1 mistake in every 20 words. Lower is better.
Is Whisper still a good choice?
Yes. It is open-weight, covers 90+ languages and remains the most-used speech model. Newer hosted models can be more accurate in English.
How much does transcription cost?
Hosted models typically cost $0.001–$0.01 per minute. Use the calculator above with your monthly minutes to compare the top models.
Can I switch models later?
Yes. With LLM API every model uses the same request format, so switching is a one-line change.
