Jev call copilot CurrencyTransfer research and architecture built 2026-09-27

nova-3 cost and timing (@cf/deepgram/nova-3 via Workers AI REST)

Measured with scripts/probe-stt.mjs (method and raw rows: docs/runbook/stt.md).

Input Audio (min) Neurons Neurons/min stt_ms Date
sample mp3, 32 kb/s 9.765 4,616.006 472.7 9,072 2026-09-25
m4a / wav / 39-min mp3 not measured (HTTP 429, daily free neuron allocation used up) 2026-09-25

472.7 neurons/min matches the price table's 472.73 neurons/min for HTTP nova-3 ($0.0052/min, $0.312/h; docs/research/stt-options.md). The workers-free plan includes 10,000 neurons/day, which is about 21 minutes of audio. A 30-40 minute call (14-19k neurons) costs more than a full day's free allocation.

Devbox faster-whisper: formats and long files (the default engine)

Measured end to end through the deployed Worker (upload → devbox job → POST /api/calls/:id/segments) on 2026-09-26 at f7d69bf. Method, commands and journal lines: docs/runbook/stt-devbox.md, section Formats and long files. The engine is large-v3-turbo int8 on CPU, 16 threads, VAD on, condition_on_previous_text off, under Nice=10. It costs nothing per minute.

Input Type Audio (s) Transcribe (s) RTF Segments Peak RSS (MB)
5-minute cut, m4a (AAC 64k) audio/mp4 300.0 86.2 0.287 82 1,668
5-minute cut, wav (16 kHz PCM) audio/wav 300.0 81.4 0.271 79 1,668
sample ×4 mp3 (39:03.5) audio/mpeg 2,343.5 702.9 0.300 678 3,266

Diarisation adds 35.4 s for each 5-minute cut and 277.9 s for the 39-minute file (RTF 0.12). Direct runs on the full-length files in the feasibility probe measured the same: m4a RTF 0.293 (177 segments), wav 0.285 (189 segments), 39-minute mp3 0.291 (682.5 s, 678 segments, 3.42 GB peak RSS). All three types stay in UPLOAD_TYPES, and WAV stays capped at 25 MB.

314 words · Internal working documents. Do not share outside CurrencyTransfer.