Moonshine Base
Free.ai (self-hosted)
·
stt
·
~500 tokens per minute
Moonshine Base is a speech-to-text model built by Useful Sensors. Strongest at Low-latency live transcription, embedded devices.. Self-hosted on Free.ai GPUs — runs free against your daily token pool (500 tokens per minute). Released under MIT — commercial use permitted on Free.ai.
Use via API
OpenAI-compatible REST API. Generate a key and call this model in seconds.
curl -X POST https://api.free.ai/v1/stt/ \
-H "Authorization: Bearer sk-free-..." \
-H "Content-Type: application/json" \
-d '{"model":"moonshine-base","audio_url":"https://..."}'Frequently Asked Questions
Moonshine Base transcribes spoken audio into text. Upload an MP3, WAV, M4A, or video file and Moonshine Base returns the full transcript plus optional SRT/VTT subtitles with timestamps.
Moonshine Base handles dozens of languages — Whisper-family models cover 90+, Parakeet covers ~25, others vary. Pick "auto-detect" or specify the language for highest accuracy.
Word-error rate is 5–10% on clean English audio, 10–20% on noisy or accented audio. Large variants of the same architecture do meaningfully better on hard cases — pick larger when the audio is rough.
Yes — every segment includes start/end timestamps. Export as SRT or VTT and the times map straight onto your video.
Moonshine Base runs on our own GPUs against your daily free pool first; $5 → 200,000 paid tokens after that. About ~500 tokens per minute.
MP3, WAV, M4A, FLAC, OGG, plus video (MP4, MOV, WebM) — we extract the audio. Max 500 MB per upload. Longer files? Split with /audio/cut/ or use /v1/stt/batch/.
Speaker diarization is a separate pass — toggle "diarize" on /transcribe/. Moonshine Base handles the transcription; diarization labels each segment with Speaker 1 / Speaker 2 / etc.
Yes — /batch/ accepts a folder of audio files. Each transcript lands in /account/?tab=history with the original filename. For folder-tree preservation use the API.
Yes — POST your audio to /v1/stt/transcribe/ with model="Moonshine Base". Returns JSON with text + segments + word-level timestamps. /api/ has the full reference.
Self-hosted models keep audio on our GPUs; premium pass through with a DPA. Audio is deleted after the share-window (24h anon, 7d signed-in). We do not train on your inputs.
Yes — Free.ai grants commercial use of transcripts. You need rights to the audio you uploaded (your own recording, licensed material, or content with consent).
Real-time factor is roughly 0.05–0.2× — a 60-minute podcast transcribes in 3–12 minutes. Premium models often finish faster. Use the queue button to close the tab.