Speech and vector capabilities, one API key
Beyond text models, ShinLinkAI offers four capabilities: text-to-speech, one-shot voice cloning, speech-to-text and text embeddings. They all run on our own compute, so inputs never leave our own hardware — suited to call QA, meeting notes, spoken content and knowledge-base Q&A.
Listen: three built-in voices
Listen: one-shot voice cloning
Four capabilities
Text-to-speech (TTS)
Natural Chinese speech synthesis with several built-in voices; long text is synthesized in segments and output as 22.05kHz WAV. Billed at 1 point per character (¥0.001/character).
Voice cloning
Zero-shot cloning: a 5–10 second reference recording is enough to replicate a voice — suited to branded narration, audiobooks and digital-human presenting. Billed at 3 points per character.
Speech-to-text (ASR)
Chinese and English speech to text with automatic sentence segmentation, suited to meeting notes, archived support calls and call QA. Billed at 5 points per second (¥0.005/second).
Text embeddings
1024 / 768-dimension vectors for knowledge-base retrieval, deduplication and clustering, paired with text models for RAG. Billed at 3 points per thousand tokens.
Measured performance (internal test data)
| Capability | Measured | Notes |
|---|---|---|
| Speech-to-text | RTF 0.55 (10.2s audio → 5.1s returned) | About 2× faster than real time; parallel streams supported |
| Text-to-speech | RTF ≈ 1.0 (44 characters → 9–10s audio, taking 9–10s) | Dual compute cards, dual channels; 3 concurrent streams finished in 13.3s |
| Voice cloning | RTF ≈ 2.5 (6.7s reference audio; 22 characters synthesized in 16.5s) | Includes reference-voice feature extraction; suited to offline batch generation |
| Embeddings (bge-m3) | 64 items/second (short text) · 45 items/second (long text) | 1024 dimensions |
| Embeddings (nomic) | 109 items/second (short text) · 82 items/second (long text) | 768 dimensions |
Integration
Speech capabilities share the same API key as text models, use standard HTTP endpoints, and return audio binary or JSON:
Billing (¥1 = 1000 points): text-to-speech 1 point/character (¥0.001/character), voice cloning 3 points/character, speech-to-text 5 points/second (¥0.005/second), embeddings 3 points/thousand tokens.
Getting started: speech capabilities share the same API key and balance as text models — no separate application, just call them after signing up. For bulk usage and enterprise cooperation, email official@shinloo.com.