Build Audio Projects With the Industry's Most Capable API

LLMs operate on text tokens — by the time audio is converted to a transcript, acoustic signals like tone, prosody, pause timing, and voice characteristics are permanently discarded. Velma's Ensemble Listening Model (ELM) processes audio natively with 100+ specialized models running in parallel, so it retains the signals that text-first systems throw away at step one. This produces outputs (emotion from voice, deepfake detection, behavioral risk scoring) that no transcript-based approach can replicate.

The API accepts .aac, .flac, .m4a, .mp3, .mp4, .ogg, .opus, .wav, and .webm files up to 50 MB per call. For longer recordings, use the streaming endpoint which processes audio in real time with sub-second latency on the first result.

Both. Every model family offers a batch (REST) endpoint for recorded files and a streaming (WebSocket) endpoint for live audio. The streaming mode returns results utterance-by-utterance as audio arrives, with most models producing a first result within 2 seconds. This is designed for real-time call analysis, live agent assist, and voice authentication flows.

Custom behaviors are defined in plain English via the API — no fine-tuning required. You describe what you want to detect (e.g. 'The agent promised a callback but did not provide a time') and Velma uses that description as a detection rule. You can also start from one of 100+ built-in behavior templates and modify them. Custom behavior support is currently in the Velma Triage tier.

Multilingual Transcription supports over 90 languages with automatic language detection. English Fast Transcription is tuned specifically for English with lower latency and higher throughput. Language Detection (in the Detect family) identifies the spoken language in under 30 seconds of audio, confidence-scored, across 100 languages.

If your project processes any kind of audio or voice — call recordings, live streams, content moderation, compliance monitoring, or voice agents — the Modulate API almost certainly covers it. Velma is the name of Modulate's core voice intelligence model; the API gives you direct access to the same intelligence that Fortune 500 companies run in production. The five model families (Detect, Transcribe, Redact, Analyze, Triage) are how that capability is packaged: each family is a standalone API endpoint, and each tier builds on the previous one. Developers typically start with the Detect or Transcribe family for a specific signal, then move up the stack as their product matures. The best way to evaluate fit is to use the free credits — 1,000 are included, no card required — and run your actual audio through the model that matches your use case.

Yes — you get 1,000 free credits to start, no card required. For Transcription specifically, that's 400 hours of audio. Rate limits in the free tier are set for development workloads. Production rate limits are configurable on paid plans, and volume discounts are available for high-throughput use cases. Contact sales for enterprise SLAs.