Introduction

The Technology Innovation Institute (TII), the Abu Dhabi lab behind the Falcon models, launched Falcon-ASR on October 6, 2026. It's a compact 1.6-billion-parameter automatic speech recognition (ASR) model that turns spoken Emirati Arabic, Modern Standard Arabic, English, French, Spanish, and Portuguese into text, with word-level timestamps. You can try it today in a free Hugging Face Space.

This is a closed release. There's no model card, no downloadable checkpoint, and no public paid API yet. The Space sends your audio to TII's own hosted deployment. Falcon-ASR shipped alongside Falcon-Emirati (a 7B dialect LLM) and Falcon-OCR-Arabic (a 270M document model), but for audio makers the speech model is the story.

Pro Tip

Start with the demo's built-in Arabic+English clip, not plain English. Code-switching between Arabic and English mid-sentence is how much of the Gulf actually talks, and it's where general-purpose ASR tends to stumble. Press play on the waveform after transcribing and each word lights up as it's spoken. That's the timestamp alignment you'd use for subtitles.

What Shipped

From TII's launch announcement and the Falcon-ASR model page:

  • Size: 1.6B parameters, small enough to pitch as a "compact" model.
  • Languages: Emirati Arabic, Modern Standard Arabic, English, French, Spanish, and Portuguese.
  • Word-level timestamps: each word comes back with its time in the recording, ready for subtitling, searchable archives, and transcript-to-video alignment.
  • Built for messy audio: TII says it's designed for background noise, overlapping speech, room reverb, and telephone-quality sound.
  • Access: a free public demo only. The weights and API stay on TII's servers.

The Numbers

TII reports two sets of results. On the Open Universal Arabic ASR Leaderboard (average across six Arabic test sets), lower is better:

ModelSizeAvg WERAvg CER

Falcon-ASR

1.6B

20.92%

8.79%

Audar-ASR-V1-Turbo

2.35B

23.17%

9.23%

Cohere Transcribe Arabic (07-2026)

2.0B

25.87%

11.80%

omniASR LLM 7B

7.0B

28.32%

12.52%

On TII's internal Emirati benchmark, Falcon-ASR scored 22.73% WER / 10.19% CER. The next-best system was Qwen3-Omni-30B-A3B-Instruct at 26.80% WER, then Audar-ASR-V1-Turbo at 27.89%, Cohere Transcribe Arabic at 31.05%, and Qwen3-ASR-1.7B at 31.52%.

Keep in mind that TII ran the Emirati test itself, on its own held-out recordings and against systems it picked. Read it as a strong sign for the dialect, not a neutral ranking.

Try It in the Demo

The Falcon ASR Demo Space needs no sign-in. Upload a file or record from your mic, then hit Transcribe. It comes with five example clips: Emirati, Saudi, MSA, English, and Arabic+English. You can copy the transcript, and mic recordings go through an adjustable noise gate (on by default at -40 dBFS).

The demo has limits: 60 seconds and 20 MB per clip, plus per-visitor and site-wide rate caps (a few runs per minute, about 20 per day). If you're testing your own podcast or voiceover audio, trim a clean 60-second slice first:

# Cut a 60s test slice starting at 1:30, mono 16 kHz WAV (what the demo sends to the API)
ffmpeg -ss 00:01:30 -t 60 -i my-episode.mp3 -ac 1 -ar 16000 -c:a pcm_s16le falcon-test.wav

Want a stress test? Record yourself reading this code-switched line at normal speed (my own script, not from TII):

Yalla, let's start the meeting. The budget review is at 3:30, بس أول شي
خلونا نراجع الـ timeline حق الـ launch, and then Sara will share the
Spanish and Portuguese subtitles. Okay? تمام.

Check whether the English terms stay in Latin script, whether "3:30" survives, and whether the timestamps stay locked as you switch languages.

Why It Matters

Most ASR leaderboards are dominated by English and Modern Standard Arabic. Emirati Arabic barely shows up in training data, so a model can get every word of an Emirati sentence technically right and still produce a transcript nobody would sign off on. A 1.6B model that tops a public Arabic leaderboard and handles code-switched speech with word timing is a real building block for Gulf-market subtitling, dubbing prep, voice agents, and TTS dataset labeling, if TII opens up weights or an API. For now, it's a strong demo to benchmark your Arabic audio against.

Conclusion

TII's Falcon-ASR is a 1.6B speech-to-text model for Emirati and Standard Arabic plus English, French, Spanish, and Portuguese. It posts a 20.92% average WER on the Open Universal Arabic ASR Leaderboard and 22.73% on TII's own Emirati test, with word-level timestamps built in. It's closed-weight, and the only way to run it today is the free 60-second Hugging Face demo.

Primary: TII launch announcement (Oct 6, 2026), Falcon-ASR & Falcon-OCR-Arabic model page, Falcon ASR Demo on Hugging Face,

.

—Anabel ♡