- Developed by the Technology Innovation Institute in Abu Dhabi.
- Features 1.6 billion parameters and supports five languages.
- Delivers low word error rates for Emirati and Gulf dialects.
- Available to try on the Hugging Face Demo Space.
In this guide
What is Falcon ASR?
Falcon ASR is a 1.6 billion parameter speech recognition model developed by the Technology Innovation Institute (TII) in Abu Dhabi. The model specializes in transcribing Arabic, with a particular focus on the Emirati dialect and other Gulf variations. It also supports transcribing four other languages using the same model weights.
- Developer
- Technology Innovation Institute (TII)
- Parameter Count
- 1.6 Billion
- Primary Focus
- Arabic (Emirati and Gulf dialects) and English
- Supported Languages
- Arabic, English, French, Spanish, Portuguese
- Base Architecture
- Falcon3-Audio
What are Falcon ASR's main capabilities?
Falcon ASR transcribes spoken Arabic, dialectal speech, and changes between languages. The training data includes background noise, overlapping speech, music, room reverberation, and telephony effects. This robust training helps Falcon ASR transcribe everyday conversations and phone calls accurately.
Falcon ASR generates word-level timestamps for all transcriptions. This feature links each transcribed word to its precise position in the audio file. Users can transcribe English, French, Spanish, and Portuguese using the same model weights without needing to set a language flag.
How does Falcon ASR perform?
Falcon ASR delivers high transcription accuracy on benchmark tests for both Arabic and English. In evaluations by the Technology Innovation Institute, Falcon ASR achieved an average word error rate (WER) of 20.92% across six Arabic test sets.
On internal Emirati speech evaluations, Falcon ASR achieved a 22.73% WER and a 10.19% character error rate (CER). Falcon ASR also transcribes English efficiently, recording a mean WER of 5.74% on the seven public English test sets of the Hugging Face Open ASR Leaderboard.
How to try Falcon ASR
Falcon ASR can be tested online through a public demonstration platform. Users can run transcriptions using their own audio files to evaluate the performance of the model.
- Navigate to the Hugging Face Demo Space for Falcon ASR.
- Upload your own audio recordings directly to the interface.
- Run the model to explore its transcription capabilities.
The Technology Innovation Institute plans to release API access and native applications for Falcon ASR in the future.
What are the limits of Falcon ASR?
Falcon ASR has some limitations regarding language support and integration options. Falcon ASR focuses primarily on Arabic dialects and a select group of European languages.
Additionally, developers cannot currently integrate Falcon ASR into their systems via an official API. Native applications are also unavailable at this stage. Users must rely on the Hugging Face Demo Space to test the technology.
- Strong performance on Emirati and Gulf Arabic dialects
- Word-level timestamps for exact audio alignment
- No language flags needed for supported languages
- No official API access is available yet
- Native applications are not yet released
- Supports only five languages
Frequently asked questions
What is Falcon ASR?
Which languages does Falcon ASR support?
Is there an API available for Falcon ASR?
How can I try Falcon ASR?
What makes Falcon ASR different from other Arabic speech models?
Sources: Hugging Face
Want your brand to be the answer AI gives?
See how ready your website is for ChatGPT, Gemini and Perplexity, free, in about ten seconds.