speech, voice and music
Audio AI API: transcribe speech and generate voices
Turn recordings into text, and text into natural voices, through the same API you use for language models. Whisper large-v3 for transcription, Qwen3-TTS for ready-made or cloned voices, and ACE-Step for music, all open source and all running on our NVIDIA B200 GPUs. Priced per minute of audio, per million characters or per song, billed in euros, and what you send is never used to train anything.
€5 of credit covers more than 55 hours of transcription or about 1,000,000 characters of speech with Qwen3-TTS Base (clone). Prices include VAT.
speech to text
Transcription: Whisper large-v3 speech-to-text API
Whisper is OpenAI's open-source transcription model, and large-v3 is its most accurate version. Send an audio file and get the text back, with timestamps if you ask for them, in 99 languages, including Spanish, Catalan, English, French, German and Portuguese. It costs €0.0015 per minute of audio: an hour of recording for €0.09, a thousand hours for €90.00.
Same endpoint as the OpenAI audio API: if your code already calls /v1/audio/transcriptions, point it at https://api.gpuflow.ai/v1 and it works.
curl
curl https://api.gpuflow.ai/v1/audio/transcriptions \ -H "Authorization: Bearer $GPUFLOW_API_KEY" \ -F [email protected] -F model=whisper-large-v3
text to speech
Text to speech: ready-made voices or your own, from €5.00 per million characters
A million characters is about 16 hours of spoken audio, or a 170,000-word audiobook. Pick by what you need: a catalogue of ready-made voices, or cloning from a sample.
The voices side by side
| Qwen3-TTS Base (clone) | Qwen3-TTS CustomVoice | |
|---|---|---|
| Provider | Alibaba Cloud | Alibaba Cloud |
| Voices | Clones your sample | Nine ready-made voices |
| Best for | Your own brand voice, consistent across content | Ready-made voices: notifications, IVR, read-aloud |
| Per million characters | 5.00 € | 19.00 € |
All of them answer on the same endpoint, /v1/audio/speech. To clone, you add the sample and its transcript to the same call.
music
Music generation: ACE-Step
Give it lyrics and a style and it composes a complete song, and it also makes covers in another style and redoes sections. It is not instant: you request it, wait, and download when ready. €0.50 per song.
use cases
What people build with audio AI in Europe
Call centres and sales teams
Transcribe every call, then summarise it with a language model on the same API. 10,000 minutes a month: €15.00 of transcription.
Accessibility and e-learning
Subtitles for video, audio versions of articles and courses, in the languages your audience speaks.
Healthcare and legal dictation
Sensitive recordings: we keep neither the audio nor the transcript, and nothing is used for training.
Voice for products
Assistants, audiobooks, IVR and notifications with a voice that is yours.
audio API questions
Frequently asked questions
Is the speech-to-text API compatible with OpenAI's?
Yes. Same endpoint and parameters; change the base URL and the key.
Which languages does Whisper support?
99 languages, with automatic detection. Spanish, Catalan, English, French, German, Portuguese and Italian among them.
How is text to speech billed?
Per million characters of input text. A million characters is roughly 16 hours of speech.
Can I clone a voice?
Yes, with Qwen3-TTS Base, from a sample of 1 to 30 seconds and its exact transcript. You are responsible for having the rights to the voice.
Is the audio I send kept?
No. Audio, transcripts and generated speech are processed in memory and returned in the same call. The exception is music: we keep the lyrics, the style and the song so you can download it. Nothing you send is used to train any model.
Try the audio API with €5 free
Activated with your card, and you only pay for what you use.