Add AI services to your application without a live room.

Translation · File transcription · WAV · Vision · Moderation

Translation

$5 per million input Unicode code points. TranslateGemma 4B and Hy-MT2 1.8B cover 45 enabled language/script entries, with up to 500 code points per request. Both languages must share an enabled model; coverage is not all-to-all and there is no English pivot. Explicitly select zh, zh-Hant or yue for Chinese script/variety. Dense inputs can exceed token bounds; truncated output is rejected without a successful-translation charge.

File transcription + diarization

$0.003 per decoded audio minute, rounded once per file to started seconds. Upload audio/video up to 256 MiB and two hours, then poll for Whisper transcription, timestamps and optional speaker labels. Language defaults to English and must be explicit; auto is not supported. Labels are anonymous and file-local. They do not identify people or guarantee overlapping-speaker separation.

Supported container families include WAV, FLAC, MP3, MP4/MOV, Matroska/WebM, Ogg and AAC, subject to decoder validation. Only the first audio stream is used. Files are privately staged and decoded in an isolated preparation container before inference. No public URL import is supported.

Text/SSML to WAV

$0.006 per 1,000 expanded spoken Unicode code points. Choose voice, language, speed, optional SSML/dictionary and watermark. Submit ≤10,000 source/expanded code points in ≤64 KiB JSON. Receive a complete mono 48 kHz PCM16 WAV and integrity metadata. The outer output bound is 30 minutes; SSML is limited to 60 seconds per document. Charge occurs only for completed valid output.

Vision snapshots

$0.002 per completed snapshot. Upload a static JPEG/PNG/WebP (≤6 MiB, ≤4 million pixels) with an optional question (≤4,000 code points). Gemma 4 E2B produces a bounded answer (up to 96 output tokens). This API does not observe WebRTC rooms, process PDFs or accept full video files.

Audio and image/frame moderation

Audio costs $0.004 per submitted minute. Send raw PCM16 mono 16 kHz chunks of at most 30 seconds. Static JPEG/PNG/WebP frames cost $0.0005 each, with the same image limits as vision. At 12 frames/minute this equals $0.006/minute; other sampling rates have different costs.

Use multipart HTTP or a ticket-authenticated WebSocket with one in-flight item per socket. Applications own capture, frame sampling and enforcement. Results are versioned combined-policy verdicts with uncalibrated native scores; errors are not pass verdicts. There is no automatic mute/kick, public text-moderation endpoint or full-video upload analysis.

Private results, bounded work

Requests use wallet signatures and prepaid credits. Inference is bounded and may return 429 when busy. Results/downloads are private and retained for 24 hours. Reuse idempotency keys after uncertain replies; file uploads return the original job without comparing repeated uploads, while WAV and other inference reject changed input. Deleting completed results does not refund completed work.

Prices and integration

Use @dtelecom/x402-client. The API reference and OpenAPI specification describe request formats, polling, cancellation, errors and limits. The SDK does not automatically fund, retry or poll.