Verba Speech API Reference

Verba Speech exposes an OpenAI-compatible REST interface alongside a low-latency WebSocket streaming protocol. Every account includes 5 free hours per month forever, followed by flat $0.10/hour per-second billing across all models.


1. Quickstart

Create an account in the Developer Console using Google or email, generate a vb_live_... API key, and point the standard OpenAI SDK at https://verbaspeech.com/v1:

from openai import OpenAI

client = OpenAI(
    api_key="vb_live_...",
    base_url="https://verbaspeech.com/v1",
)

tx = client.audio.transcriptions.create(
    model="verba/speech-finance-v1",
    file=open("call.m4a", "rb"),
    prompt="NVDA, ASML, Colette Kress, non-GAAP",
)
print(tx.text)

2. Authentication

Pass your API key in the standard HTTP Authorization header on all REST requests:

Authorization: Bearer vb_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxx

3. Models & Pricing

Model ID Mode Best For Free Tier Pay-As-You-Go
verba/speech-v1 Batch REST General meetings, media, call centers, noisy audio 5 hrs/mo shared $0.10 / hr
verba/speech-finance-v1 Batch REST Earnings calls, tickers, analyst Q&A, GAAP/bps 5 hrs/mo shared $0.10 / hr
verba/speech-medical-v1 Batch REST Clinical notes, telehealth, Rx dosages, lab metrics 5 hrs/mo shared $0.10 / hr
verba/speech-realtime-v1 WebSocket Voice AI agents, live call copilots (<240ms partials) 5 hrs/mo shared $0.10 / hr

4. Batch Transcription (POST /v1/audio/transcriptions)

Accepts multipart/form-data uploads up to 100 MB per request in WAV, MP3, M4A, WEBM, OPUS, OGG, FLAC, or MP4 container formats.

5. Real-Time Streaming (WSS /v1/realtime)

Open a persistent TLS WebSocket specifying ?domain=general|finance|medical and stream 16 kHz mono PCM16 frames (20–100 ms per frame). Verba emits transcript.partial events in under 240 ms and transcript.final upon neural VAD turn completion.

6. Domain Vocabulary Biasing (prompt)

Unlike generic post-hoc spellcheckers, Verba injects your prompt terms into acoustic-linguistic decoding—achieving 99.0% executive/analyst name accuracy and 98.2% clinical pharmacology recall on unseen audio.

7. Stateless Zero Data Retention

All requests to verba/speech-medical-v1—or any request carrying the header X-Verba-Zero-Retention: true—process audio strictly in ephemeral memory. Audio buffers and transcript payloads are never written to persistent storage or used for model training.