# Models
URL: /stt/models
Learn about latest models, changelog, and deprecations.
Soniox Speech-to-Text AI provides multiple models for real-time and asynchronous
transcription and translation. This page lists the currently available models,
their capabilities, and important updates.
***
## Current models
{/*TABLE START */}
{/* NOTE: Width is set so that we have maximum of 2 lines in 'Example' column. */}
{/* NOTE: Font size is set so the table doesn't look "too big".*/}
|
Model
| {" "}
Type
| Status |
| -------------------------------------------- | ----------------------------------------------- | ------------------------------------------------ |
| **stt-rt-v6** | Real-time | **Active** |
| **stt-async-v6** | Async | **Active** |
| **stt-rt-v5** | Real-time | **Deprecated**
Will be removed Dec 15, 2026 |
| **stt-async-v5** | Async | **Deprecated**
Will be removed Dec 15, 2026 |
{/* TABLE END */}
***
## Aliases
Aliases provide a stable reference so you don't need to change your code when newer versions are released.
| Alias | Points to | Notes |
| ---------------- | -------------- | ----- |
| **stt-rt-v4** | `stt-rt-v5` | |
| **stt-async-v4** | `stt-async-v5` | |
***
## Changelog
### November 3, 2026
**New models:** stt-rt-v6, stt-async-v6
**Replaces:** stt-rt-v5, stt-async-v5
#### Overview
**Soniox STT v6** is the latest generation of Soniox speech-to-text, with major improvements for voice agents,
voice typing and meeting transcription, across 60+ supported languages.
It's available as **stt-rt-v6** for real-time streaming and **stt-async-v6** for asynchronous transcription.
V6 introduces two connected features for voice agents. With agent text input, the model follows the whole conversation,
not just the user's side. With agent voice echo cancellation, it filters the agent's own voice out of the transcript.
V6 also delivers significantly higher accuracy, gets more out of the context you provide, and brings stronger
multilingual recognition, speaker diarization and real-time translation.
#### New features
Both features are available on stt-rt-v6, in every supported language:
* **Agent text input** for much higher accuracy in voice agents. Send what your agent says on the same WebSocket
as the user's audio, and the model transcribes each reply knowing what the user is answering.
It helps most with short and ambiguous answers.
Enable it with `enable_agent_text_input`. See [agent text input](/stt/rt/agent-text-input).
* **Agent voice echo cancellation** for fewer false interruptions. With agent text input enabled, the model tells the agent's voice
apart from the user's and filters it out of the transcript, while still listening to the user, even when they interrupt.
It handles what acoustic echo cancellation leaves behind, so keep that on.
See [how it works](/stt/rt/agent-text-input#the-agents-voice-isnt-transcribed-as-the-user).
#### Key improvements
* More effective use of [context](/stt/concepts/context), especially in voice agents: context covers your business domain,
and [agent text input](/stt/rt/agent-text-input) covers what the agent says during the call
* Significantly higher transcription accuracy across every supported language
* Faster voice typing and dictation, with better recognition of numbers, currencies, units, dates, IDs, names and specialized terms
* Reliable output of spoken punctuation and formatting commands in voice typing and dictation
(commands like "period", "new line" or "scratch that" are written as words, so your app can apply them)
* Improved meeting transcription and speaker diarization for AI note taking, including long conversations and multi-speaker scenarios,
for both live meetings and recordings
* Stronger recognition of accents, multilingual conversations and language switching within a sentence
* More accurate real-time translation that adjusts how much speech it gathers before translating
#### Formatting changes
Soniox STT v6 writes prices and alphanumeric codes closer to how they're spoken, without extra formatting:
* "15 dollars and 20 cents" is written as spoken (with currency as a word), not as "$15.20".
* Hyphens in codes and IDs are written only when they're spoken, so you don't get extra hyphens like "NP-38-BX".
#### API compatibility
The v6 models are fully compatible with the existing v5 models and Soniox API.
To upgrade, simply replace the model name in your API request:
* `{ "model": "stt-rt-v6" }` for real-time
* `{ "model": "stt-async-v6" }` for async
#### Deprecation notice
The stt-rt-v5 and stt-async-v5 models will be removed on December 15, 2026.
After December 15, 2026, requests using stt-rt-v5 will automatically route to stt-rt-v6, and requests using stt-async-v5
will automatically route to stt-async-v6, with no service interruption and no API changes required.
### June 16, 2026
**New models:** stt-rt-v5
**Replaces:** stt-rt-v4
#### Overview
stt-rt-v5 is the new Soniox real-time speech-to-text model for live audio.
It delivers higher accuracy, reinvented speaker separation, improved spoken language identification, higher-quality real-time translation, faster semantic endpointing, better context handling,
and more reliable recognition and formatting of structured speech data such as numbers, dates, emails, names, addresses, and codes.
#### Key improvements
* Higher real-time transcription accuracy across 60+ languages
* Better robustness on noisy audio, telephony, far-field microphones, accents, interruptions, overlapping speech, and mixed-language conversations
* Reinvented speaker separation for identifying who said what in live conversations
* Improved spoken language identification across multilingual and accented speech
* Higher-quality real-time translation across 3,600+ language pairs
* Faster and more reliable semantic endpointing for voice agents, dictation, command systems, and conversational apps
* Better alphanumeric recognition and formatting for numbers, dates, times, emails, IDs, codes, names, and addresses
* More robust context usage for names, domain terms, product names, translation preferences, and custom vocabulary
#### API compatibility
The stt-rt-v5 model is fully compatible with the existing stt-rt-v4 model and Soniox Real-Time API.
To upgrade, simply replace the model name in your API request: `{ "model": "stt-rt-v5" }`
#### Deprecation notice
The stt-rt-v4 model will be removed on June 30, 2026.
After June 30, 2026, requests using stt-rt-v4 will automatically route to stt-rt-v5 with no service interruption and no API changes required.
### June 11, 2026
**New models:** stt-async-v5
**Replaces:** stt-async-v4
#### Overview
**stt-async-v5** is the new Soniox async speech-to-text model for processing recorded audio.
It delivers higher accuracy, stronger speaker separation, improved language identification, better context handling,
and more reliable formatting of structured speech data such as numbers, dates, emails, names, and codes.
#### Key improvements
* Higher transcription accuracy across 60+ languages
* Better robustness on noisy audio, telephony, accents, and mixed-language speech
* Completely reengineered speaker separation for identifying who said what
* Improved spoken language identification across multilingual conversations
* Better alphanumeric recognition and formatting for numbers, dates, times, emails, IDs, and codes
* More robust context usage for names, domain terms, product names, and custom vocabulary
#### API compatibility
* The stt-async-v5 model is fully compatible with the existing stt-async-v4 model and Soniox API
* To upgrade, simply replace the model name in your API request to `{ "model": "stt-async-v5" }`
#### Deprecation notice
* The stt-async-v4 model will be removed on June 30, 2026
* After June 30, 2026, requests using stt-async-v4 will automatically route to stt-async-v5 with no service interruption and no API changes required
### February 5, 2026
**New models:** stt-rt-v4
**Replaces:** stt-rt-v3
#### Overview
**Soniox v4 Real-Time** is a next-generation real-time speech recognition model built for low-latency voice interactions.
It delivers speaker-native accuracy across 60+ languages with improved latency, reliability, and conversational behavior.
The model is production-ready and fully backward-compatible with v3 Real-Time.
#### Key improvements
* Higher accuracy across all supported languages
* Better multilingual detection and mid-sentence language switching
* Lower endpoint latency with faster final transcription
* Improved semantic endpointing for more natural turn-taking
* Lower manual finalization latency with faster final transcription
* More stable, higher-quality transcription on long and multi-hour recordings
* Stronger use of provided context for domain-specific accuracy
* More fluent, accurate, and consistent translation across all supported languages
* Added `max_endpoint_delay_ms` for controlling end-of-speech endpoint delay
#### API compatibility
* The stt-rt-v4 model is fully compatible with the existing stt-rt-v3 model and Soniox API
* To upgrade, simply replace the model name in your API request:
* `{ "model": "stt-rt-v4" }` for real-time
#### Deprecation notice
* The stt-rt-v3 model will be removed on February 28, 2026
* After February 28, 2026, requests will automatically route to stt-rt-v4 with no service interruption. No API changes required
### January 29, 2026
**New models:** stt-async-v4
**Replaces:** stt-async-v3
#### Overview
**Soniox v4 Async** is the latest generation of Soniox’s asynchronous speech recognition and translation model. This release delivers a significant improvement in accuracy, robustness, and multilingual performance across more than 60 languages. v4 Async reaches human-parity transcription quality in real-world scenarios, while also introducing stronger long-form processing, improved speaker diarization, richer context handling, and higher-quality translation output. The model is designed for production-scale workloads and consistent, high-fidelity results across diverse acoustic environments and language mixes.
#### Key improvements
* Higher transcription accuracy across all languages, reaching speaker-native quality in many domains
* More robust performance in noise, accents, overlapping speech, and poor audio
* Better language identification and smoother mid-sentence language switching
* Improved speaker separation and more consistent labeling in multi-speaker audio
* Better normalization of dates, numbers, phone/email addresses, and other structured content
* More stable, higher-quality transcription on long and multi-hour recordings
* Stronger use of provided context for domain-specific accuracy
* More fluent, accurate, and consistent translation across all supported languages
#### API compatibility
* The stt-async-v4 model is fully compatible with the existing stt-async-v3 model and Soniox API
* To upgrade, simply replace the model name in your API request:
* `{ "model": "stt-async-v4" }` for async
#### Deprecation notice
* The stt-async-v3 model will be removed on February 28, 2026
* After February 28, 2026, requests will automatically route to stt-async-v4 with no service interruption. No API changes required
### October 31, 2025
#### Model retirement and upgrade
We have accelerated the retirement of older models following the overwhelmingly positive response to the new v3 models. The following models have been retired:
* stt-async-preview-v1
* stt-rt-preview-v2
Both models have been **aliased to the new Soniox v3 models.**
This means all existing requests using the old model names are now automatically served with v3, giving every user our most accurate, capable, and intelligent voice AI experience, without any code changes required.
#### Context compatibility
The context feature is now backward compatible with v3 models, ensuring smooth migration from older versions. However, we **strongly recommend updating to the new context** structure for best results and future flexibility. Learn more about [context](/stt/concepts/context).
### October 29, 2025
**Model update:** v3 enhancements
**Applies to:** stt-rt-v3, stt-async-v3
#### New features
* **Extended audio duration support:** both real-time (stt-rt-v3) and asynchronous (stt-async-v3) models now support **audio up to 5 hours** in a single request.
#### Quality improvements
* **Higher transcription accuracy** across challenging audio conditions and diverse languages.
#### Notes
* No API changes are required; existing integrations continue to work seamlessly.
* For asynchronous processing, large files up to 5 hours can now be uploaded directly without chunking.
* For real-time streaming, sessions up to 5 hours are supported under the same WebSocket connection.
### October 21, 2025
**New models:** stt-rt-v3, stt-async-v3
**Replaces:** stt-rt-preview-v2, stt-async-preview-v1
#### Overview
The **v3 models** introduce major improvements across recognition, translation, and reasoning — making Soniox faster, more accurate, and more capable than ever before.
These models power real-time and asynchronous speech processing in 60+ languages, with enhanced accuracy, robustness, and context understanding.
#### Key improvements
* Higher transcription accuracy across 60+ languages
* Improved multilingual switching — seamless recognition when speakers change language mid-sentence
* Significantly higher translation quality, especially for languages such as German and Korean
* The async model now also supports translation
* Support for new advanced structured context, enabling richer domain- and task-specific adaptation
* Enhanced alphanumeric accuracy (addresses, IDs, codes, serials)
* More accurate speaker diarization, even in overlapping speech
* Extended maximum audio duration to 5 hours for both async and real-time models
#### API compatibility
* The v3 models are fully compatible with the existing Soniox API, if you are not using the context feature.
* To upgrade, simply replace the model name in your API request:
* `{ "model": "stt-rt-v3" }` for real-time
* `{ "model": "stt-async-v3" }` for async
* If you are using the context feature, update to the new structured [context](/stt/concepts/context) for improved accuracy.
#### Deprecation notice
The following preview models are **deprecated** and will be retired on **November 30, 2025:**
* stt-async-preview-v1
* stt-rt-preview-v2
Please migrate to the v3 models before that date to ensure uninterrupted service.
### August 15, 2025
* Deprecated `stt-rt-preview-v1`
### August 5, 2025
* Released `stt-rt-preview-v2`
* Higher transcription accuracy
* Improved translation quality
* Expanded to support all translation pairs
* More reliable automatic language switching
* **Replaces:** stt-rt-preview-v1