The update also improves transcription accuracy, speaker diarization, language detection, and endpoint detection across 60+ languages.
Two New Features
Agent Text Input
STT v6 receives the text spoken by an AI agent as additional recognition context. This helps interpret short or ambiguous user responses based on the preceding conversation.
Agent Voice Echo Cancellation
The model filters the agent’s audio playback from incoming speech while continuing to recognize the user. This enables interruptions without incorrectly treating the agent’s voice as user input.
Both features target fewer false interruptions and repeated clarification requests.
Technical Specifications
| Specification | Soniox STT v6 |
| Model type | Automatic speech recognition (ASR) |
| Languages | 60+ |
| Architecture | Unified multilingual model |
| Processing | Real-time streaming, asynchronous |
| Agent text input | Supported |
| Agent voice echo cancellation | Supported |
| Speaker diarization | Improved |
| Language identification | Improved |
| Endpoint detection | Improved |
| Context customization | Enhanced |
| Mixed-language recognition | Supported |
| Real-time translation | Supported |
Soniox has not disclosed v6 parameter counts, word error rates (WER), or detailed latency benchmarks.
Soniox STT v6 vs. STT v5
| Feature | STT v5 | STT v6 |
| Languages | 60+ | 60+ |
| Real-time transcription | Yes | Yes |
| Agent text input | Not documented | Added |
| Agent voice echo cancellation | Not documented | Added |
| Speaker diarization | Supported | Improved |
| Language identification | Supported | Improved |
| Endpoint detection | Supported | Improved |
| Context customization | Supported | Enhanced |
| Multilingual translation | Supported | Supported |
STT v5 Performance Baseline
Soniox's August 2026 Pipecat benchmark reported:
- Semantic WER: 1.34%
- Perfect transcripts: 83.3%
- Median finalization latency: 260 ms
- P95 finalization latency: 305 ms
- Test size: 1,000 speech samples
Equivalent v6 benchmark results have not been confirmed.
Availability
Soniox states that STT v6 is available for:
- Real-time speech recognition and voice agents.
- Asynchronous transcription.
- Multilingual meeting transcription.
- Voice typing and real-time translation.
Exact v6 API model identifiers, pricing, and migration requirements were not established in the announcement.
Conclusion
STT v6 builds on STT v5 with agent-aware recognition and echo cancellation, allowing voice agents to process user speech during their own playback.
The release also introduces claimed accuracy improvements across 60+ languages. However, without published v6 WER and latency measurements, the performance gains over v5 cannot yet be quantified.
