Skip to main content
Tinfoils technical security guarantees make transcribing or voicing sensitive content (medical notes, legal drafts, internal communications) safe in a way no conventional cloud audio API can match. File transcription and text-to-speech are OpenAI-compatible, so existing clients work by changing only the base URL and key. Realtime transcription is only partially OpenAI-compatible (details below). See Audio models for the full list of models.

​
Realtime transcription

Voxtral Mini Realtime (voxtral-mini-4b-realtime) streams speech-to-text over a WebSocket: you send PCM16 audio chunks as the user speaks, and partial transcripts stream back word-by-word.
This is only available through the tinfoil-js WebSocket right now. Contact us if you need it in another SDK.
wss://inference.tinfoil.sh/v1/realtime
Authenticate with your API key as a Bearer token in the Authorization header. The tinfoil-js SDK handles this for you and pins the TLS connection to the attested enclave key: 
import { TinfoilAI } from "tinfoil";

const client = new TinfoilAI({ apiKey: "<YOUR_API_KEY>" });
const rt = await client.realtime({ model: "voxtral-mini-4b-realtime" });

rt.on("session.created", (event) => console.log(event));
rt.send({ type: "input_audio_buffer.append", audio: base64AudioChunk });
This is Node.js only: browsers can’t expose TLS certificate details, so the connection can’t be pinned to the attested key, and verified realtime isn’t available in the browser yet. If this is a problem, contact us.

​
OpenAI Realtime-compatible mode

Connect with ?intent=transcription and the endpoint speaks the OpenAI Realtime transcription dialect. Existing OpenAI Realtime clients should work by changing only the URL and key: 
wss://inference.tinfoil.sh/v1/realtime?intent=transcription
The session flow:
  1. The server sends session.created.
  2. Optionally send session.update declaring your input format (defaults to PCM16 mono at 24kHz; declare {"type": "audio/pcm", "rate": 16000} if you capture at 16kHz). The server replies session.updated.
  3. Stream audio with input_audio_buffer.append (base64 PCM16). Partial transcripts arrive as conversation.item.input_audio_transcription.delta events.
  4. Send input_audio_buffer.commit to end the utterance. The server replies with input_audio_buffer.committed and a conversation.item.input_audio_transcription.completed event carrying the final transcript. You can then stream the next utterance on the same connection.
import { TinfoilAI } from "tinfoil";

const client = new TinfoilAI({ apiKey: process.env.TINFOIL_API_KEY! });
const rt = await client.realtime({ model: "voxtral-mini-4b-realtime" });

// Partial transcripts stream in as deltas; the final transcript arrives on completion.
rt.on("conversation.item.input_audio_transcription.delta", (event) => {
  process.stdout.write(event.delta);
});
rt.on("conversation.item.input_audio_transcription.completed", (event) => {
  console.log("\nFinal:", event.transcript);
});

// Stream PCM16 audio as base64 chunks, then commit to end the utterance.
for (const chunk of pcm16Chunks) {
  rt.send({ type: "input_audio_buffer.append", audio: chunk.toString("base64") });
}
rt.send({ type: "input_audio_buffer.commit" });
Turn Detection: This only works in push to talk dictation mode. The openAI spec uses server_vad turn detection: the server detects silence, and emits a speech_started/stopped event. Our mode does not do that. See more in the realtime model repo

​
Native mode

Connecting with ?model=voxtral-mini-4b-realtime instead speaks the leaner vLLM dialect: flat transcription.delta events, a session.update carrying only {"model": ...}, client commits with a final flag, and transcription.done (with usage) after the final commit. Audio must be PCM16 mono at 16kHz. One transcription session per connection.

​
File transcription

For recorded audio on disk, transcribe the whole file in one request over the OpenAI-compatible /v1/audio/transcriptions endpoint. The model is required: use voxtral-small-24b for accuracy or whisper-large-v3-turbo for fast, lightweight transcription. 
from tinfoil import TinfoilAI

client = TinfoilAI(api_key="<YOUR_API_KEY>")

with open("meeting.mp3", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="voxtral-small-24b",  # or "whisper-large-v3-turbo"
        file=audio,
    )

print(result.text)

​
Text-to-speech

Synthesize speech over the OpenAI-compatible /v1/audio/speech endpoint, which returns WAV audio. Use qwen3-tts for low-latency speech or voxtral-tts for the larger multilingual model. Both require a voice. 
from tinfoil import TinfoilAI

client = TinfoilAI(api_key="<YOUR_API_KEY>")

response = client.audio.speech.create(
    model="qwen3-tts",
    voice="serena",
    input="Your audio never leaves the enclave.",
)
with open("speech.wav", "wb") as f:
    f.write(response.read())
Voices are the upstream models’ default preset voices. qwen3-tts has 9: aiden, dylan, eric, ono_anna, ryan, serena, sohee, uncle_fu, vivian. voxtral-tts has 20: neutral_female, neutral_male, casual_female, casual_male, cheerful_female, ar_male, de_female, de_male, es_female, es_male, fr_female, fr_male, hi_female, hi_male, it_female, it_male, nl_female, nl_male, pt_female, pt_male.