TII Releases Falcon-ASR, a 1.6B Arabic Speech Model With Self-Reported 20.92% WER

The model uses one set of weights for all five languages, does not require a language flag and outputs text in the language spoken, according to the report. It also returns word-level timestamps, a feature useful for subtitling, searchable recordings and meeting notes. TII has not disclosed the model's maximum audio context, run environment, architecture details or training data size; the hosted demo accepts clips up to 60 seconds and 20 MB.

Access is limited. Falcon-ASR currently runs only as a hosted demo on Hugging Face, with an API and native apps planned. No weights or quantized variants have been published, and no license is disclosed. TII's benchmark figures are self-reported, and Falcon-ASR's Arabic average is TII's own run, not an entry in the Open Universal Arabic ASR Leaderboard run by the ELM Research Center. TII followed that leaderboard's protocol and pinned manifests, checking competitor averages on September 30, 2026.

In the Arabic comparisons, Falcon-ASR's 20.92% average WER and 8.79% average character error rate place it ahead of Audar-ASR-V1-Turbo (2.35B parameters, 23.17% WER and 9.23% CER), Cohere Transcribe Arabic from July 2026 (2.0B, 25.87% WER and 11.80% CER), and omniASR LLM 7B (7.0B, 28.32% WER and 12.52% CER). The Audar model card lists its own average CER as 9.20%, slightly lower than TII's reported figure for Audar. Lower error rates are better.

TII also built an internal Emirati and Gulf test set from held-out recordings with human-validated transcripts, complementing the UAE subset of the public Casablanca dataset. On that private set, Falcon-ASR recorded 22.73% WER and 10.19% CER, ahead of Qwen3-Omni-30B-A3B-Instruct at 26.80% WER and 12.72% CER, Audar-ASR-V1-Turbo at 27.89% WER and 13.75% CER, Cohere Transcribe Arabic at 31.05% WER and 18.07% CER, Qwen3-ASR-1.7B at 31.52% WER and 13.35% CER, and Audar-ASR-V1-Flash at 32.87% WER and 15.36% CER. The test set is private, so the result cannot be reproduced; Audar reports 19.4% WER on a different Emirati set called Mixat, so the numbers are not directly comparable.

For English, TII reports a 5.74% mean WER across the seven test sets used by the Hugging Face Open ASR Leaderboard. Per-set scores range from 1.75% on LibriSpeech clean to 11.86% on Earnings-22, which the report identifies as its weakest English test set. Qwen3-ASR-1.7B lists 5.76% mean WER on the same leaderboard. TII says Falcon-ASR also supports French, Spanish and Portuguese but has published no scores for those languages.

Falcon-ASR builds on TII's Falcon3-Audio work, which describes audio-language models built on instruction-tuned large language models and Whisper encoders and trained on fewer than 30,000 hours of public audio. The exact architecture and training data size of Falcon-ASR have not been published, according to MarkTechPost. The model's closest Arabic-focused alternatives differ in openness: Audar-ASR-V1-Turbo is listed with the AudarAI Community License, Cohere Transcribe Arabic under Apache 2.0, and Qwen3-ASR-1.7B under Apache 2.0. Falcon-ASR's license is not disclosed.

In dialectal Arabic, transcribed resources are far scarcer than for Modern Standard Arabic. TII trained Falcon-ASR on Emirati, MSA, other Gulf and Arabic dialects, and English, targeting everyday speech including dialect forms and language switching. Training audio was augmented with background noise, overlapping speech, music, room reverberation and telephony effects, and speed and pitch were varied. The same treatment was applied to Emirati recordings.

The release gives Arabic voice product developers a compact model that leads larger systems in TII's own comparisons, but the practical barrier is access. Without published weights, a license or an API, developers cannot yet deploy Falcon-ASR outside the hosted demo. The scores TII reports are internally produced, and the Emirati test set is not public.