VoicePasteModel compatibility
Recognition and polishing
These tables reflect the built-in configuration. Each recognition model has its own row. Polishing services and local runtimes are listed individually; choose a specific language model within your service. An integration listed here does not mean every account, model version and parameter combination has been tested live.
| Model | Runs | Highlights | Native streaming | Hotwords | Setup guide |
|---|---|---|---|---|---|
| Doubao Streaming ASR 2.0 | Online | Chinese–English speech; dialect support varies by API mode | Yes | Yes | Get API key |
| ElevenLabs Scribe v2 Realtime | Online | Accurate multilingual transcription with automatic language detection | Yes | Yes | Get API key |
| SenseVoice INT8 | Local | Compact quantized model for Chinese, English, Japanese, Korean and Cantonese | No | No | Local setup |
| FunASR-Nano INT8 | Local | Chinese, English and Japanese; Chinese dialects and singing voice | No | Yes | Local setup |
| Zipformer Chinese–English | Local | Native streaming, mixed Chinese–English speech and low latency | Yes | Yes | Local setup |
| Zipformer Chinese INT8 | Local | Native streaming with a compact, quantized Chinese model | Yes | Yes | Local setup |
| Qwen3-ASR 0.6B INT8 | Local | Multilingual speech, Chinese dialects and singing voice | No | Yes | Local setup |
| Service / runtime | Runs | Setup guide |
|---|---|---|
| DeepSeek | Online | Get API key |
| OpenAI | Online | Configure service |
| Gemini | Online | Configure service |
| Anthropic | Online | Configure service |
| 硅基流动 / SiliconFlow | Online | Configure service |
| OpenRouter | Online | Configure service |
| Ollama | Local | Configure service |
| Custom OpenAI-compatible service | Online / Local | Configure service |
Understanding recognition features
- Native streaming means the current VoicePaste integration receives ongoing recognition results. Both Zipformer variants stream locally.
- Simulated preview is available for SenseVoice, FunASR-Nano and Qwen3-ASR through the app’s offline recognizer integration. It is distinct from native streaming.
- Hotwords require support from both the model and its configuration. Enable the relevant keyterm setting for ElevenLabs. VAD and punctuation models are supporting components, rather than separate transcription services.
- Doubao dialects depend on the API mode and model version. The official documentation distinguishes bidirectional streaming from streaming input; capabilities vary between modes. Configure the service using its API documentation.
Online and local
Online recognition sends audio to your chosen service. Online polishing sends the text to be processed. To keep both on your device, use local recognition with local polishing, or turn polishing off.
Ollama runs local models; install a model and enable its compatible endpoint. SiliconFlow and OpenRouter are model platforms: use your own API key and a model name provided by the platform. Custom services must expose an OpenAI-compatible Chat Completions endpoint; arbitrary APIs are not interchangeable. The Anthropic preset uses its compatibility endpoint and requires supported models and parameters.
Configure recognition · Configure polishing and templates
Official capability references
These references describe the upstream models and APIs. VoicePaste’s current integration may expose a subset of those capabilities.