Ultra

AI model · Speech to text · Audio / speech → Text by superwhisper

High-accuracy cloud speech recognition for challenging audio and accents.

A legacy cloud-based speech recognition model for Superwhisper Pro users, optimized for difficult audio, heavy accents, and specialized vocabulary across 100+ languages.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Launched
2026-08-06

Category: Natural Language Processing

Modalities
Audio → Text
Published accuracy
10.6% Word Error Rate (average across 8 datasets)
Released
Updated September 3, 2026
Languages
100+ languages
Word Error Rate (WER)
10.6%
Processing Speed
24x real-time
  • 10.6% Word Error Rate (WER) average
  • 24x real-time processing speed
  • 0.49s average response latency
  • Supports 100+ languages and dialects
  • Optimized for heavy accents and overlapping speech

Overview

Ultra is a legacy cloud-based speech recognition model developed by Superwhisper for professional users requiring high accuracy in challenging audio conditions. It was specifically engineered to handle non-American accents, overlapping speech, and specialized vocabulary that on-device models often struggle to process. The model operates at 24x real-time speed with an average response latency of 0.49 seconds and supports over 100 languages. While it served as the primary cloud engine for Superwhisper Pro, it has since been superseded by the faster and more accurate S1-Voice model. Ultra remains available as a record of performance benchmarks and for users who have not yet transitioned to the newer S1 family of models.

Transcription Capabilities

  • Challenging Accents: The model was specifically built to process audio that on-device models find difficult, including accents far from American English.
  • Overlapping Speech Support: It is designed to handle complex audio environments where multiple people are talking over each other.
  • Specialized Vocabulary Recognition: The model accurately transcribes specialized names and terms that general-purpose models may not recognize.
  • Multilingual Support: It supports transcription for more than 100 different languages and dialects.

Performance and Delivery

  • 24x Processing Speed: The model operates at 24x real-time speed, meaning a one-minute voice note is processed in approximately three seconds.
  • 0.49s Response Latency: It delivers text results with a response latency of approximately 0.49 seconds.
  • 10.6% Word Error Rate: The model achieves a word error rate of 10.6% as measured across eight standardized datasets.
  • Cloud-Based Processing: Audio is sent to Superwhisper servers for processing rather than running locally on the user's device.

Benchmarks

  • AMI Meetings: 15.6% WER; Common Voice Australian English: 4.4% WER; LibriSpeech Other: 3.4% WER; Loquacious: 9.1% WER

Use cases

  • Professional teams transcribe multi-party meetings with overlapping speech and far-field microphones for accurate records.
  • Developers use the model to recognize specialized technical names and jargon that standard on-device models might miss.
  • Global users dictate in over 100 different languages and dialects with high accuracy for international communication.
  • Users with heavy non-American accents utilize the cloud-based processing to achieve lower word error rates than local models.

Specifications

Modalities
Audio → Text
Published accuracy
10.6% Word Error Rate (average across 8 datasets)
Access
Available through the Superwhisper Pro desktop and mobile applications via cloud proxy
Released
Updated September 3, 2026
Languages
100+ languages
Word Error Rate (WER)
10.6%
Processing Speed
24x real-time
Response Latency
0.49s
Language Support
100+ languages
Deployment Type
Cloud-based
Blended Benchmark Score
73

How it works

  1. Model Selection: The user selects the Ultra model within the Superwhisper application settings on macOS, Windows, iOS, or Android.
  2. Audio Transmission: The user records audio or uploads a file; the audio is then sent to the Superwhisper cloud proxy for processing.
  3. Cloud Transcription: The model transcribes the audio into text, handling accents and specialized vocabulary, and returns the result in approximately 0.49 seconds.

Pricing

Superwhisper Pro (Monthly)
$8.49 per month
Superwhisper Pro (Annual)
$84.99 per year
Superwhisper Pro (Lifetime)
$249.99 one-time payment

FAQ

Is Superwhisper Ultra accurate enough for meetings?

Ultra was built for difficult audio scenarios, including multi-party meetings with overlapping speech and far-field microphones. It achieved a 15.6% word error rate on the AMI Meetings dataset.

Why is speed shown as a multiple rather than seconds?

Speed is shown as a multiple of real time because it scales with audio length; at 10x real time, a one-minute voice note takes about six seconds to process.

Is Ultra the latest model available?

Superwhisper Ultra has been superseded by the S1-Voice model, which is faster and more accurate. The company provides Ultra's benchmarks as a record of historical performance.

Added by AntSpace from public sources (superwhisper.com).

Related

More AI models by superwhisper