venice-audio-transcription（威尼斯音频转录）

Venice Transcription (/audio/transcriptions)

POST /api/v1/audio/transcriptions takes an audio file and returns text. It's OpenAI-compatible with multipart/form-data — the OpenAI SDK's audio.transcriptions.create() works unchanged.

Use when

• You need STT (speech-to-text) for voice notes, meetings, podcasts, short audio.
• You need timestamps for subtitles / chapters.
• You want to pick between fast local-style models (Parakeet) and large multilingual ones (Whisper, Wizper, Scribe).

For long video / YouTube transcription, see venice-video's /video/transcriptions (takes a public video URL directly).

Minimal request

curl https://api.venice.ai/api/v1/audio/transcriptions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -F "file=@./meeting.m4a" \
  -F "model=nvidia/parakeet-tdt-0.6b-v3" \
  -F "response_format=json" \
  -F "timestamps=false"

{ "text": "Alright everyone, let's kick off the meeting..." }

With timestamps=true, json format also returns segment/word timings (schema is model-specific).

Request (multipart/form-data)

Models

GET /models?type=asr returns the current catalog. ASR pricing is pricing.per_audio_second.usd — cost scales with audio duration.

OpenAI SDK

import OpenAI from 'openai'
import fs from 'node:fs'

const client = new OpenAI({
  apiKey: process.env.VENICE_API_KEY,
  baseURL: 'https://api.venice.ai/api/v1',
})

const out = await client.audio.transcriptions.create({
  file: fs.createReadStream('meeting.m4a'),
  model: 'openai/whisper-large-v3',
  response_format: 'json',
  language: 'en',
  // @ts-expect-error — Venice-specific extra, passes through multipart
  timestamps: true,
})

console.log(out.text)

Batch / long files

Venice doesn't expose native chunking. For files > ~30 min, split client-side on silence with ffmpeg or pydub, transcribe each chunk, then concatenate with offset timestamps.

ffmpeg -i long.mp3 -f segment -segment_time 600 -c copy chunk_%03d.mp3

Errors

Gotchas

• file must be uploaded as a real multipart file part. JSON + base64 is not supported here.
• Timestamps are only surfaced in the JSON response shapes (json, verbose_json, srt, vtt). With response_format: text the handler returns a plain text/plain body containing just the transcript — you'll lose any timestamp data, so pick verbose_json / srt / vtt when you need timings.
• language is Whisper-specific. Parakeet / Scribe ignore it and auto-detect.
• Peak concurrency limits apply — on 429, back off; big batches should throttle to ~5 parallel requests.
• Content-policy rejection on the transcript is returned as 422 with an error string; it does not surface suggested_prompt on this path.