Use Inference APIs with VoiceInk

Use Inference APIs in VoiceInk for cloud transcription and for the AI enhancement step, through VoiceInk's custom model settings. One key covers both, billed per audio minute and per token with no subscription.

Settings

Setting	Value
Base URL	`https://api.inferenceapis.com/v1` (also `/openai/v1`)
API key	From API Keys; send as `Authorization: Bearer …`
Chat model ids	`openai/gpt-oss-120b`, `deepseek-ai/DeepSeek-V4-Flash`, `zai-org/GLM-5.3-Flash`, `meta-llama/Llama-3.3-70B-Instruct-Turbo` — all models
Audio model ids	`openai/whisper-large-v3` (transcription), `hexgrad/Kokoro-82M` (speech)

Which model

Transcription: openai/whisper-large-v3 for any language, or nvidia/parakeet-tdt-0.6b-v3 if you dictate in English or another European language and want the fastest result. Enhancement: meta-llama/Llama-3.3-70B-Instruct-Turbo. Check accuracy for your language first. Prices and measured speed: chat models.

Steps

In VoiceInk open AI Models and find Custom Transcription Models. Choose Add transcription model.
Display Name: anything, for example Inference APIs Whisper.
API Endpoint: https://api.inferenceapis.com/v1/audio/transcriptions. VoiceInk wants the full URL, not just the base.
API Key: your Inference APIs key.
Model Name: openai/whisper-large-v3. Switch Multilingual Model on. For Parakeet use nvidia/parakeet-tdt-0.6b-v3.
Press Test, then select the model for dictation.
For enhancement, under Custom Enhancement Models choose Add enhancement model with endpoint https://api.inferenceapis.com/v1/chat/completions, the same key, and model meta-llama/Llama-3.3-70B-Instruct-Turbo.

Configuration

Check the endpoint from a terminal first

curl -s https://api.inferenceapis.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $INFERENCE_API_KEY" \
  -F file=@recording.wav -F model=openai/whisper-large-v3 -F response_format=json

Verify

Smoke test

Press Test on the model card; VoiceInk reports "Test successful". Then dictate one sentence.

How this guide was checked

Configuration only We do not have a Mac to run VoiceInk on. The field names below are read from VoiceInk's source (CustomProviderManagementView.swift), and on 2026-09-18 we sent this API the exact request VoiceInk's OpenAICompatibleTranscriptionService builds: a multipart upload of a 16 kHz WAV with model, response_format=json, temperature=0 and language. A 7.8-second clip came back in 1.3 seconds on Whisper Large v3 and 0.8 seconds on Parakeet. The enhancement request shape (OpenAI chat completions) was checked the same way. If a step does not match what you see, tell us.

Gotchas

The endpoint field takes the full path ending in /audio/transcriptions (or /chat/completions for enhancement). The base URL alone returns a JSON 404.
DeepSeek, GLM and GPT-OSS models think before they answer, which adds seconds to every dictation. Llama 3.3 70B does not, so it is the safe choice for enhancement unless VoiceInk lets you send reasoning_effort "none".
Whisper Large v3 costs $0.002 per audio minute here: an hour of actual speaking is 12 cents. Silence you do not record is not billed.
Audio is processed on GPUs in the United States and not kept after the response. VoiceInk's local models remain the choice if nothing may leave the Mac.

If something fails

401 — key missing or wrong · 402 insufficient_balance — add credits on Billing · 404 model_not_found — check the id against the model list (aliases such as gpt-oss-120b work too) · 503 model_unavailable — the model is being enabled. Full details in the API docs.