Speech to text for a fraction of a cent.
Live or one-shot transcription in Spanish, Catalan, Galician, Basque, English, Portuguese, French, Italian, German, for $0.084 an hour, billed per second. The audio is transcribed on our own servers in the EU.
Spoken words, written down.
Speech to text, or automatic speech recognition, turns audio of people talking into text. A voice agent needs it to understand the caller; on its own it transcribes calls, meetings, voice notes and dictation.
Ours is our own model, Conv 2.0: one recogniser for nine languages, on our own GPUs, with no audio sent to a third-party cloud. A typical utterance is transcribed in about 55 ms, and the text comes back written out: digits, phone numbers, e-mail addresses and punctuation.
- LiveWebSocket streaming with automatic utterance detection, for transcripts while people talk.
- Who said whatSpeaker labels: every word comes back with its timing and its speaker, from an API compatible with ElevenLabs Scribe (
POST /v1/speech-to-text). In the dashboard it is the “Speaker labels” option in File mode. API reference. - Speaker labels cost$0.0017 per audio minute on top of speech to text, with no minimum per request.
- One-shotSend a recording over REST, get the text and timed segments back. Any sample rate.
- LanguagesSpanish, Catalan, Galician, Basque, English, Portuguese, French, Italian, German, with one model.
- Price$0.0014 a minute, $0.084 an hour, billed per second from a prepaid balance.
One request.
curl -X POST "https://api.lokutor.com/stt/transcribe" \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "audio=@recording.wav" \
-F "language=es" For live transcription, use WS /ws/stt or the SDKs.
Common questions.
How much does Lokutor’s speech to text cost?
$0.0014 a minute ($0.084 an hour), billed per second from a prepaid balance with no plan. New accounts get $5 of credit with no card.
Which languages does it transcribe?
Spanish, Catalan, Galician, Basque, English, Portuguese, French, Italian, German: every language our voice speaks. Send the language with the request when you know it; without it the language is detected.
Which model does Lokutor use for speech recognition?
Our own, Conv 2.0, in production since 20 October 2026 on our own GPUs in the EU; before that we served NVIDIA’s open Parakeet TDT 0.6B. As served today it scores 6.05% weighted word error on English conversational speech, scored like Artificial Analysis’s, and 9.62% across its eight other languages (Catalan, German, Spanish, Basque, French, Galician, Italian and Portuguese); with its language model, not yet in production, 5.82% and 7.56%. The method and the full table are on the benchmarks page.
Can it tell who is speaking?
Yes. With speaker labels on, each word comes back with its timing and a speaker. On one real 60-minute meeting with five English speakers it found all five and attributed 99.4% of the words to the right speaker (8.0% diarization error). It misses short interjections and overlapping speech, and we have not yet measured it on phone calls or on audio in other languages. Speaker detection uses pyannote, credited at lokutor.com/credits/.
How much do speaker labels cost?
$0.0017 per audio minute, on top of the speech-to-text price, with no minimum per request.
How long do speaker labels take?
Speaker labels run on their own GPU, which starts on demand. The first request after a quiet spell takes about 8 minutes while the card starts; later requests answer in seconds. An hour of audio is processed in about 2.3 minutes, and files can be up to 2 hours long.
Is my audio sent to another company?
No. Transcription runs on our own servers; the audio is not sent to a third-party speech service.
Transcribe your first file.
There is a speech-to-text studio in the dashboard. $5 of credit to start.