BizCrush

BizCrush API

Build with realtime multilingual translation and live speech recognition. Stream audio, get live captions, and reach audiences in their own language.

Automatic language detection across 60 languages — no configuration required. Just send audio and get results.

Also available: file STT, text-to-speech, and text translation. MCP Guide →

Beta This API is currently in beta. Endpoints and behavior may change without prior notice.

Realtime Translation

All plans

One audio stream. Live captions and translations in the languages you choose.

Quick start

  1. 1

    Connect with your API key

    Create a key in Settings, then open this WebSocket. Replace YOUR_API_KEY with your key.

    wss://extapi.bizcrush.ai/v1/translate/stream?api_key=YOUR_API_KEY
  2. 2

    Choose your languages

    Send this JSON as the first text frame. This example translates English speech into Spanish. Omit language_hints for automatic input-language detection.

    Choose 1–10 output languages in target_languages.

    {
      "target_languages": ["es"],
      "language_hints": ["en"],
      "enable_diarization": true
    }
  3. 3

    Wait for the connection, then send audio

    After {"connected":true}, send audio as binary frames at the speed it is spoken.

    PCM16 (default) or Opus

    PCM16: 16 kHz, mono, little-endian; maximum 32,000 bytes per frame.

    Sending Opus audio

    Use this as your first config for browser WebM/Opus. Send one audio-only recording, including its initial headers. Blocks may span multiple WebSocket messages.

    {
      "encoding": "opus",
      "container": "webm",
      "target_languages": ["es"],
      "language_hints": ["en"]
    }

    For raw Opus, use "container":"raw" and send one complete packet per binary message. Use "container":"ogg" for Ogg/Opus. Mono and stereo are accepted and converted to 16 kHz mono. WebM must contain exactly one Opus audio track.

    Send 20–100 ms chunks at realtime speed. Each message may contain at most 64 KiB and one second of decoded audio. Omit container to auto-detect Ogg or WebM headers. Keep glossary and other options in the same config.

  4. 4

    Read the captions and finish the stream

    You receive the original text, translations, and speaker labels. Update each caption by its id; translations can arrive after the original text is final.

    When audio ends, send an empty binary frame or {"type":"finish"}. Keep reading until {"finished":true}, then let the server close.

Browser microphone: WebM/Opus

Keep every dataavailable chunk in order, including the last chunk after stop(). Do not restart the recorder or resend its header on the same WebSocket.

const mimeType = "audio/webm;codecs=opus";
if (!MediaRecorder.isTypeSupported(mimeType)) throw new Error("Use PCM16 on this browser");
const stream = await navigator.mediaDevices.getUserMedia({audio: true});
const recorder = new MediaRecorder(stream, {mimeType});
const ws = new WebSocket("wss://extapi.bizcrush.ai/v1/translate/stream?api_key=YOUR_API_KEY");
let pending = Promise.resolve();
recorder.ondataavailable = ({data}) => {
  if (data.size) pending = pending.then(async () => {
    const bytes = await data.arrayBuffer();
    if (ws.readyState === WebSocket.OPEN) ws.send(bytes);
  });
};
recorder.onstop = async () => {
  await pending;
  if (ws.readyState === WebSocket.OPEN) ws.send(JSON.stringify({type: "finish"}));
  stream.getTracks().forEach(track => track.stop());
};
ws.onopen = () => ws.send(JSON.stringify({"encoding": "opus", "container": "webm", "language_hints": ["en"], "target_languages": ["es"]}));
ws.onmessage = ({data}) => {
  const event = JSON.parse(data);
  if (event.connected) recorder.start(100);
  console.log(event);
};
ws.onclose = () => {
  if (recorder.state !== "inactive") recorder.stop();
  stream.getTracks().forEach(track => track.stop());
};
// Call recorder.stop() when the user finishes.

Custom translations & synonyms

Help recognition with important names, and choose how your terms should appear in each translated language.

Recognition hints

List names and domain terms in context_keywords to help the recognizer.

Preferred translations

Map a source phrase to your preferred wording with translation_terms, separately for each output language.

To give synonyms or spelling variants the same translation, map each source phrase to the same target:

customer support/customer serviceatención al cliente(es)
source
Source phrase, synonym, or spelling variant
target
Preferred wording in the translation
target_language
Output language code for this mapping

For custom translations, use this config as the first frame in step 2. Send one config per connection.

{
  "target_languages": ["es"],
  "language_hints": ["en"],
  "translation_terms": [
    {"source": "customer support", "target": "atención al cliente", "target_language": "es"},
    {"source": "customer service", "target": "atención al cliente", "target_language": "es"}
  ]
}
See a glossary with multiple output languages

Send this complete config once, before audio. Each entry needs source, target, and an output target_language. The language may be omitted when there is only one output. Mappings guide model wording; source captions retain the recognized text and exact translations can vary with context.

{
  "target_languages": [
    "es",
    "fr"
  ],
  "language_hints": [
    "en"
  ],
  "context_keywords": [
    "customer support",
    "customer service"
  ],
  "translation_terms": [
    {
      "source": "customer support",
      "target": "atención al cliente",
      "target_language": "es"
    },
    {
      "source": "customer service",
      "target": "atención al cliente",
      "target_language": "es"
    },
    {
      "source": "customer support",
      "target": "service client",
      "target_language": "fr"
    },
    {
      "source": "customer service",
      "target": "service client",
      "target_language": "fr"
    }
  ],
  "enable_diarization": true
}
Response format & late translation updates

Merge translation_update.translations into the caption with the same id, by language. Keep previously received languages when applying a patch. Treat missing or null translations as an empty map. Events are connected, chunk, translation_update, pong, finished, or error.

{"chunk":{"id":"caption-1","text":"Hello, how can I help you?","language":"en","speaker":"0","is_final":true,"translations":{}}}
{"translation_update":{"id":"caption-1","translations":{"es":"Hola, ¿en qué puedo ayudarte?"}}}
Other authentication methods

Server clients can use X-API-Key. Firebase ID tokens are accepted as Authorization: Bearer FIREBASE_ID_TOKEN or ?authorization=FIREBASE_ID_TOKEN. These speech APIs use Firebase ID tokens; MCP uses OAuth access tokens. Authentication failures return HTTP 401; an unavailable backend returns HTTP 503.

Limits & token usage
LimitValue
Output languages1–10 per connection
Input-language hintsUp to 3
Keywords / glossary entriesUp to 100 each
Combined context5,900 characters, including normalized text and separators
Audio frameUp to 32,000 bytes
Session lengthUp to 55 minutes

Send the first audio within 30 seconds; pings do not extend this deadline. During silence, send {"type":"ping"} at least every 30 seconds. Audio may run at most one second ahead of elapsed session time.

Minimum charge: one minute. For N output languages, the first valid audio prepays 100 + 60 × N tokens. After the first 60 seconds, each started 30-second period prepays 50 + 30 × N tokens.

Custom context can add a one-time startup charge: each output includes 100 UTF-8 bytes; each started 20 bytes above that costs 1 token. The full reference explains exactly how keywords and glossary bytes are counted.

Payment must succeed before translation starts. Insufficient tokens or payment failures stop the stream. No audio means no charge or provider session.

At capacity, new connections return HTTP 503 with Retry-After. Reconnect with exponential backoff and jitter. Existing streams keep their capacity; a stalled upstream ends with an error. Always wait for finished=true to confirm completion.

Available on every plan. Usage is token-based, with a one-minute minimum. Meetings and transcripts are not stored automatically.

Building with an agent? Full agent reference · OpenAPI JSON

Get Your API Key

Create up to 5 API keys from your account settings for speech and text translation.

Get API Key

Speech-to-Text API

Featured

Speech-to-Text API for audio transcription with speaker diarization. Supports file-based (batch) and real-time (streaming) transcription.

Build Your Own STT Service

Use the BizCrush STT API to build your own speech-to-text service. Integrate real-time transcription into your apps, process audio files at scale, or create custom voice-powered workflows — all with speaker diarization and multi-language support.

Automatic language detection across 60 languages — no configuration required. Just send audio and get results.

Overview

Base URL https://extapi.bizcrush.ai/v1

Authentication

STT, TTS, and realtime translation require authentication. Send your API key in the X-API-Key header or use this query parameter, including for browser WebSockets:

?api_key=YOUR_API_KEY

You can issue your API key at bizcru.sh/settings.

Firebase ID tokens are also accepted as Authorization: Bearer FIREBASE_ID_TOKEN or ?authorization=FIREBASE_ID_TOKEN. These speech endpoints use Firebase ID tokens; MCP uses OAuth access tokens.

For agents and code generators: full API reference · OpenAPI JSON.

Live STT (WebSocket)

Real-time audio streaming transcription via WebSocket. Send audio chunks and receive interim/final results in real time.

WS wss://extapi.bizcrush.ai/v1/stt/stream

Query Parameters

Name Type Required Description
api_key string one option API key, required when no authentication header or Firebase ID token is provided.
format string optional "json" for JSON text frames, omit for protobuf binary
language_hints string[] optional Language codes as comma-separated (?language_hints=ko,en) or repeated (?language_hints=ko&language_hints=en). Up to 3; extras dropped. Omit for auto-detection.
enable_diarization string optional Speaker diarization toggle. "false" to disable; enabled by default. When enabled, each chunk carries a speaker label.

Protocol

1 Send Config

Use ?format=json for the JSON protocol and examples below. Send a JSON text frame with encoding and optional session_id. If format is omitted, send protobuf SessionStreamTranscribeRequest binary messages for config, every audio chunk, and the empty end-of-audio marker; responses are SessionStreamTranscribeResponse.

Field Type Description
encoding string "pcm16" or "opus" (default: "pcm16")
session_id string Session identifier (optional, auto-generated if omitted)
language_hints string[] Optional language codes, e.g. ["ko","en"] or "ko,en". Up to 3 entries; extras dropped. Overrides query param. Omit for auto-detection.
enable_diarization boolean Speaker diarization toggle (default: true). Overrides the query param.
context_keywords string[] Domain keywords/terms to bias recognition toward names, jargon, etc., e.g. ["Albert Gee","BizCrush"] or "Albert Gee,BizCrush". Up to 2000 characters total; extras dropped.
{"encoding": "pcm16", "session_id": "optional-session-id"}
2 Receive Connection Status

Server responds with connection status

{"connected": true}
3 Stream Audio

Send raw audio as binary WebSocket frames

Encoding Format Chunk Size
pcm16 16kHz, mono, 16-bit little-endian 640 bytes (20ms) recommended
opus Raw Opus (container: "raw"), Ogg/Opus or WebM/Opus Opus frame unit
4 Receive Transcription Results

Server sends interim and final results as JSON text frames

Field Type Description
chunk.id string Chunk ID — same across interim/final for one utterance, new ID after final
chunk.session_id string Session ID
chunk.text string Transcribed text (cumulative — each interim contains full text so far)
chunk.is_final boolean false: interim result, true: final confirmed result
chunk.speaker string Speaker label ("0", "1", "2", ...) when diarization is enabled. Omitted when disabled or the speaker is unresolved.
{"chunk": {"id": "01KMCHP57H31XYZABC", "session_id": "my-session", "text": "Hello, how are you?", "is_final": true, "speaker": "1"}}
5 Signal End-of-Audio (required for clean shutdown)

When no more audio will be sent, send an empty binary frame (0 bytes) and KEEP THE WEBSOCKET OPEN. The server finalizes any pending utterances (including ones still in interim state), streams the trailing is_final results back, then closes the WebSocket. Closing the WebSocket without this marker drops the trailing finals — WebSocket has no half-close, so once the client closes the server cannot deliver the remaining results. Finalize completes within ~5 seconds after the EOS marker. After the trailing finals the server sends a terminal frame ({"finished": true} for JSON, disconnect_reason=SESSION_ENDED for protobuf) and then closes.

ws.send(b'')                  # Python
ws.send(new ArrayBuffer(0));  // JavaScript

Examples

Python

import asyncio
import json
import websockets

async def live_stt(api_key: str):
    url = f"wss://extapi.bizcrush.ai/v1/stt/stream?api_key={api_key}&format=json"

    async with websockets.connect(url) as ws:
        await ws.send(json.dumps({"encoding": "pcm16"}))

        resp = json.loads(await ws.recv())
        assert resp["connected"], f"Connection failed: {resp}"

        async def send_audio():
            with open("audio.pcm", "rb") as f:
                while chunk := f.read(640):
                    await ws.send(chunk)
                    await asyncio.sleep(0.02)
            # End-of-audio marker. DO NOT close the WebSocket here — the
            # server needs the connection open to deliver trailing finals.
            # The server will close after finalize completes (~5 seconds).
            await ws.send(b'')

        async def receive_results():
            try:
                async for msg in ws:
                    data = json.loads(msg)
                    if "chunk" in data:
                        chunk = data["chunk"]
                        status = "FINAL" if chunk["is_final"] else "interim"
                        speaker = chunk.get("speaker", "?")
                        print(f"[{status}] (S{speaker}) {chunk['text']}")
            except websockets.exceptions.ConnectionClosed:
                pass

        await asyncio.gather(send_audio(), receive_results())

asyncio.run(live_stt("YOUR_API_KEY"))

JavaScript (Browser)

const ws = new WebSocket(
  "wss://extapi.bizcrush.ai/v1/stt/stream?api_key=YOUR_API_KEY&format=json"
);

ws.onopen = () => {
  ws.send(JSON.stringify({ encoding: "pcm16" }));
};

ws.onmessage = (event) => {
  const data = JSON.parse(event.data);

  if (data.connected) {
    console.log("Connected! Start sending audio...");
  }

  if (data.chunk) {
    const { id, text, is_final, speaker } = data.chunk;
    console.log(`[${is_final ? "FINAL" : "interim"}] (S${speaker ?? "?"}) ${text}`);
  }
};

// Server closes the WebSocket after delivering trailing finals.
ws.onclose = () => {
  console.log("Stream finalized by server");
};

function sendAudioChunk(pcmData) {
  if (ws.readyState === WebSocket.OPEN) {
    ws.send(pcmData);
  }
}

// Call this when you have no more audio to send.
// DO NOT call ws.close() — the server needs the socket open to deliver
// trailing finals, then closes it itself (~5 seconds).
function endStream() {
  if (ws.readyState === WebSocket.OPEN) {
    ws.send(new ArrayBuffer(0));  // end-of-audio marker
  }
}
Browser microphone: WebM/Opus

Keep every dataavailable chunk in order, including the last chunk after stop(). Do not restart the recorder or resend its header on the same WebSocket.

const mimeType = "audio/webm;codecs=opus";
if (!MediaRecorder.isTypeSupported(mimeType)) throw new Error("Use PCM16 on this browser");
const stream = await navigator.mediaDevices.getUserMedia({audio: true});
const recorder = new MediaRecorder(stream, {mimeType});
const ws = new WebSocket("wss://extapi.bizcrush.ai/v1/stt/stream?format=json&api_key=YOUR_API_KEY");
let pending = Promise.resolve();
recorder.ondataavailable = ({data}) => {
  if (data.size) pending = pending.then(async () => {
    const bytes = await data.arrayBuffer();
    if (ws.readyState === WebSocket.OPEN) ws.send(bytes);
  });
};
recorder.onstop = async () => {
  await pending;
  if (ws.readyState === WebSocket.OPEN) ws.send(new ArrayBuffer(0));
  stream.getTracks().forEach(track => track.stop());
};
ws.onopen = () => ws.send(JSON.stringify({"encoding": "opus", "container": "webm", "language_hints": ["en"]}));
ws.onmessage = ({data}) => {
  const event = JSON.parse(data);
  if (event.connected) recorder.start(100);
  console.log(event);
};
ws.onclose = () => {
  if (recorder.state !== "inactive") recorder.stop();
  stream.getTracks().forEach(track => track.stop());
};
// Call recorder.stop() when the user finishes.

File STT

Send an audio file URL and receive the full transcript with speaker diarization.

Complete WebM/Opus recordings are supported. Assemble all MediaRecorder chunks into one file and provide a downloadable URL; later chunks alone do not contain the required header.

POST /stt Auth: ?api_key

Transcribe an audio file URL and return the full transcript with speaker diarization.

Request Body (JSON)

Name Type Required Description
audio_url string required Public audio file URL (signed URL, CDN URL, etc.)
session_id string optional Session identifier. Auto-generated if omitted
enable_diarization boolean optional Enable speaker diarization (default: true)
language_hints string[] optional Language codes, e.g. ["ko","en"] or "ko,en". Up to 3; extras dropped. Omit for auto-detection. Legacy `language_hint` (single string) also accepted.
context_keywords string[] optional Domain keywords/terms to bias recognition toward names, jargon, etc., e.g. ["Albert Gee","BizCrush"] or "Albert Gee,BizCrush". Up to 2000 characters total; extras dropped.

Response Fields

Field Type Description
text string Full transcript text
detected_language string Detected language code
confidence float Average confidence (0~1)
utterances array Speaker-diarized utterance segments
utterances[].speaker string Speaker number ("0", "1", "2", ...)
utterances[].text string Utterance text
utterances[].start_ms integer Start time (milliseconds)
utterances[].end_ms integer End time (milliseconds)
utterances[].confidence float Utterance confidence
utterances[].language string Utterance language

Response Example

{
  "text": "Could you introduce yourself and tell us about your role?",
  "detected_language": "en",
  "confidence": 0.961,
  "utterances": [
    {
      "speaker": "1",
      "text": "Could you introduce yourself and tell us about your role?",
      "start_ms": 960,
      "end_ms": 5400,
      "confidence": 0.94,
      "language": "en"
    },
    {
      "speaker": "2",
      "text": "Sure. My name is Ethan Kim, and I'm CTO of BizCrush.",
      "start_ms": 5820,
      "end_ms": 9600,
      "confidence": 0.96,
      "language": "en"
    }
  ]
}

cURL Example

curl -X POST \
  "https://extapi.bizcrush.ai/v1/stt?api_key=YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"audio_url": "https://example.com/audio.oga", "enable_diarization": true}'

Note: Response time depends on audio length — may take up to several minutes. Set timeout to 600+ seconds.

Text-to-Speech (TTS)

Convert text to natural-sounding speech and receive a signed download URL (audio uploaded to Firebase Storage).

POST /tts Auth: ?api_key

Generate speech audio from text and return a signed download URL (audio uploaded to Firebase Storage at tts/{session_id}/{message_id}.ogg).

Request Body (Protobuf)

application/x-protobuf (request and response are protobuf-encoded SessionTTSRequest / SessionTTSResponse)

Name Type Required Description
session_id string required Session identifier
message_id string required Message identifier (used in storage path)
text string required Text to convert to speech

Response Fields

Field Type Description
audio_url string Signed download URL (Opus / OGG) — valid for several days
error string Set on failure; audio_url empty
Audio Format
OGG/Opus, single channel
Default Voice
nova (configurable server-side)

TTS Supported Languages

Text-to-speech generates natural-sounding audio across 57 languages. The language is detected automatically from the input text — no configuration required.

Code Language
en English
zh 中文
hi हिन्दी
es Español
ar العربية
fr Français
pt Português
ru Русский
id Bahasa Indonesia
de Deutsch
ja 日本語
vi Tiếng Việt
it Italiano
ko 한국어
th ภาษาไทย
af Afrikaans
hy Հայերեն
az Azərbaycan
be Беларуская
bs Bosanski
bg Български
ca Català
hr Hrvatski
cs Čeština
da Dansk
nl Nederlands
et Eesti
fi Suomi
gl Galego
el Ελληνικά
he עברית
hu Magyar
is Íslenska
kn ಕನ್ನಡ
kk Қазақ
lv Latviešu
lt Lietuvių
mk Македонски
ms Bahasa Melayu
mi Te Reo Māori
mr मराठी
ne नेपाली
no Norsk
fa فارسی
pl Polski
ro Română
sr Српски
sk Slovenčina
sl Slovenščina
sw Kiswahili
sv Svenska
tl Tagalog
ta தமிழ்
tr Türkçe
uk Українська
ur اردو
cy Cymraeg

Supported Languages

Automatic language detection is built in — no configuration needed. The STT engine detects and transcribes across 60 languages seamlessly, even when speakers switch languages mid-conversation. Optionally use language_hint for faster detection.

Code Language
en English
zh 中文
hi हिन्दी
es Español
ar العربية
fr Français
pt Português
ru Русский
id Bahasa Indonesia
de Deutsch
ja 日本語
vi Tiếng Việt
it Italiano
ko 한국어
th ภาษาไทย
af Afrikaans
sq Shqip
az Azərbaycan
eu Euskara
be Беларуская
bn বাংলা
bs Bosanski
bg Български
ca Català
hr Hrvatski
cs Čeština
da Dansk
nl Nederlands
et Eesti
fi Suomi
gl Galego
el Ελληνικά
gu ગુજરાતી
he עברית
hu Magyar
kn ಕನ್ನಡ
kk Қазақ
lv Latviešu
lt Lietuvių
mk Македонски
ms Bahasa Melayu
ml മലയാളം
mr मराठी
no Norsk
fa فارسی
pl Polski
pa ਪੰਜਾਬੀ
ro Română
sr Српски
sk Slovenčina
sl Slovenščina
sw Kiswahili
sv Svenska
tl Tagalog
ta தமிழ்
te తెలుగు
tr Türkçe
uk Українська
ur اردو
cy Cymraeg

STT Error Codes

Status Error Description
401 Missing authorization API key not provided
401 Invalid API key Invalid API key
400 Empty request body Request body is empty
400 Missing audio_url audio_url field missing (File STT)
400 Missing session_id, message_id, or text Required TTS field missing
400 Invalid JSON JSON parsing failed
500 TTS generation failed OpenAI TTS call failed
500 Storage upload failed Firebase Storage upload failed
500 (varies) Internal server error (transcription failure, etc.)
WS 1008 Auth failure WebSocket closed with code 1008 on authentication failure

Text Translation

Translate plain text into multiple target languages with BizCrush AI.

POST /v1/translate-text

Translate plain-text input while preserving document formatting.

Parameters

Name Type Required Description
text str required Plain text to translate (up to 100,000 characters by default)
target_languages list[str] required Target language codes or names, such as ["ko", "en", "Japanese"] (up to 10)
source_language str optional Source language code or name (default: auto)