This document describes the available Speech-to-Text (STT) cloud providers, their purposes, and official documentation.
AriaType supports multiple cloud STT providers to transcribe audio into text. Each provider has different characteristics in terms of latency, accuracy, cost, and language support.
| Provider | Type | Latency | Best For | Languages |
|---|---|---|---|---|
| Volcengine Streaming | WebSocket Real-time | Low (~200ms) | Live dictation, real-time transcription | 60+ languages |
| Volcengine Flash | HTTP Batch | Medium | Short recordings, batch processing | 60+ languages |
| OpenAI Whisper | HTTP Batch | Medium | High accuracy transcription | 50+ languages |
| OpenAI Realtime | WebSocket Real-time | Low | Real-time with GPT-4o capabilities | Multiple |
| Deepgram | WebSocket Real-time | Very Low (~300ms) | Fast streaming, cost-effective | 30+ languages |
| Custom Endpoint | HTTP/WebSocket | Varies | Self-hosted or custom STT services | Varies |
Real-time streaming STT via WebSocket. Provides partial results during speech, ideal for live dictation scenarios.
- Real-time voice input
- Live transcription with immediate feedback
- Continuous speech recognition
wss://openspeech.bytedance.com/api/v3/sauc/bigmodel_nostream
- App ID: Application identifier from Volcengine Console
- Access Token: Authentication token (may expire, needs refresh)
- Resource ID (optional): Default
volc.bigasr.sauc.duration
- Partial results during speech
- Language auto-detection
- Punctuation and formatting
- Low latency (~200ms chunk processing)
STT should output raw, unprocessed transcription. The Polish engine handles formatting (ITN, punctuation, deduplication). STT's job is accurate phonetic transcription, not polished text.
Volcengine configuration:
{
"enable_itn": true, // Numbers: "一九七八" → "1978" — Polish engine handles formatting
"enable_punc": true, // Punctuation: "今天天气好" → "今天天气好。" — Polish engine handles formatting
"enable_ddc": false // Disfluency/deduplication — OFF: preserve raw output
}HTTP-based batch STT for short audio files. Lower cost than streaming, suitable for offline processing.
- Short audio transcription (< 60 seconds)
- Batch processing of recordings
- Cost-sensitive scenarios
https://openspeech.bytedance.com/api/v3/auc/bigmodel/recognize/flash
- App ID: Application identifier
- Access Token: Authentication token
- Resource ID (optional): Model identifier
- Lower cost than streaming
- No real-time results
- Suitable for batch processing
- Good for short recordings
OpenAI's Whisper model via batch API. High accuracy transcription with support for 50+ languages.
- High-accuracy transcription
- Multi-language support
- Scenarios where latency is not critical
https://api.openai.com/v1/audio/transcriptions
- API Key: OpenAI API key from platform.openai.com
- Model (optional): Default
whisper-1
- High accuracy
- 50+ language support
- Translation capability (to English)
- Price: $0.006/minute
OpenAI's Realtime API with GPT-4o for low-latency speech-to-text with advanced capabilities.
- Real-time transcription with AI capabilities
- Voice-to-voice applications
- Advanced audio understanding
wss://api.openai.com/v1/realtime
- API Key: OpenAI API key with Realtime API access
- Model:
gpt-4o-realtime-preview-2026-12-17
- Very low latency
- GPT-4o powered understanding
- Function calling support
- Audio output capability
Deepgram's streaming STT via WebSocket. Fast, accurate, and cost-effective for real-time transcription.
- Fast streaming transcription
- Cost-effective real-time STT
- High-volume applications
wss://api.deepgram.com/v1/listen
- API Key: Deepgram API key from console.deepgram.com
- Model (optional): Default
nova-2 - Language (optional): Language code (e.g.,
en-US,zh-CN)
- Very fast (~300ms latency)
- Nova-2 model for high accuracy
- Interim results
- Smart formatting and punctuation
- Cost-effective pricing
OpenAI-compatible custom STT endpoint for self-hosted or third-party STT services.
- Self-hosted STT services
- Third-party STT providers with OpenAI-compatible API
- Custom STT implementations
- API Key: Authentication key for your service
- Base URL: Your STT API endpoint
- Model (optional): Model identifier
- Self-hosted Whisper models
- Azure Speech Services
- Google Cloud Speech-to-Text
- Custom STT implementations
{
"provider_type": "custom",
"api_key": "your-api-key",
"base_url": "https://your-stt-service.com/v1/audio/transcriptions",
"model": "whisper-large-v3"
}- You need real-time transcription for live dictation
- Low latency is critical
- You're primarily targeting Chinese users
- You have short recordings (< 60 seconds)
- Cost is a primary concern
- Real-time results are not needed
- Accuracy is the top priority
- You need multi-language support
- Latency is acceptable
- You need advanced AI capabilities
- Voice-to-voice interaction is required
- Budget allows for premium service
- You need fast, cost-effective streaming STT
- High volume with good accuracy
- English-language focus
- You have self-hosted STT infrastructure
- You need a provider not directly supported
- You want full control over STT pipeline