Microsoft AI’s MAI-Transcribe-2 undercuts OpenAI, Google and ElevenLabs on price and speed
Microsoft AI on Thursday released MAI-Transcribe-2, a speech-recognition model the company claims is faster, more accurate, and cheaper than those offered by OpenAI, Google, or ElevenLabs. Priced at just 10 cents per hour of audio, this represents a significant reduction from the initial pricing of $0.36 per hour five months ago. For enterprises processing 100,000 hours of call-center audio annually, the new price drops the bill from $36,000 to $10,000.
This release marks a strategic shift for Microsoft, as it builds its own cutting-edge models and integrates them into products that previously relied on OpenAI's technology. The focus is on transcription, with MAI-Transcribe-2 demonstrating the company's progress.
What MAI-Transcribe-2 offers:
- 60 languages: An upgrade from 43 languages in June's MAI-Transcribe-1.5 and a vast improvement over the original release's 25 languages.
- Advanced Features: Built on Microsoft Foundry and MAI Playground, the model handles background noise, low-quality recordings, and overlapping speech, making it suitable for real-world business scenarios.
- Comprehensive Transcription:
- Speaker Diarization: Distinguishes between speakers in multi-person recordings.
- Word-Level Timestamps: Enables search, editing, and alignment with video.
- Keyword Biasing: Allows developers to train the model on specific domains, reducing misidentification of jargon.
- Automatic Language Identification: Removes the need for users to specify languages in advance.
- Customizable Output: Offers "verbatim" mode for preserving all filler words and "clean" mode for readable captions, catering to compliance, legal, and other specialized needs.
- Code Switching: Handles conversations switching between languages mid-sentence, supporting markets like Hinglish and Spanglish.
Performance Benchmarks:
Microsoft makes three performance claims, each based on different benchmarks:
-
MAI-Transcribe-2 tops the FLEURS benchmark with a 5.2% average word error rate across 60 languages. FLEURS, developed by Google, uses native speakers reading sentences in 102 languages to establish a standard for multilingual speech recognition.
-
Microsoft also references its own internal benchmarks, claiming superior performance in terms of speed and accuracy.
-
The company's claims should be evaluated considering the context and limitations of each benchmark.