Audio modelLimited access

Falcon-ASR

Technology Innovation Institute

Released: October 6, 2026
Data date: October 7, 2026

Model class: Automatic speech recognition
Access: Public Hugging Face demo linked by TII; this does not establish weight or API access

Falcon-ASR transcribes Emirati Arabic, Modern Standard Arabic, English, French, Spanish, and Portuguese. The 1.6-billion-parameter ASR returns word-level timestamps, which also support subtitles and searchable recordings.

The model can be used through TII's public Hugging Face demo. For Arabic documents, TII also offers Falcon-OCR-Arabic.

Specifications and access

Specification	Value and source
Model class	Automatic speech recognitionSource
Input modalities	Audio containing spoken Emirati Arabic, Modern Standard Arabic, English, French, Spanish, or PortugueseSource
Output modalities	Transcribed text with word-level timestampsSource
Access	Public Hugging Face demo linked by TII; this does not establish weight or API accessSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

English ASR mean WER across seven test sets (Show measurement, test conditions, and source)

Source value: 5.74
Score: 5.74
Metric: mean word error rate across seven test sets
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. TII's mean across LibriSpeech clean/other, SPGISpeech, VoxPopuli, GigaSpeech, AMI, and Earnings-22; per-set weights and sample counts are not published.
Context: This is the vendor's mean; the source does not state whether aggregation is size- or audio-weighted.

Source TII Falcon-ASR article

Open Universal Arabic ASR Leaderboard mean WER (Show measurement, test conditions, and source)

Source value: 20.92
Score: 20.92
Metric: word error rate
Unit: %
Benchmark version: Leaderboard manifests checked 2026-09-30
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B; six public Arabic test sets under the Open Universal Arabic ASR Leaderboard protocol, equally weighted across test sets; lower WER/CER is better.
Context: TII does not publish per-set clip counts or the six individual Arabic test-set scores in this article.

Source TII Falcon-ASR article

Open Universal Arabic ASR Leaderboard mean CER (Show measurement, test conditions, and source)

Source value: 8.79
Score: 8.79
Metric: character error rate
Unit: %
Benchmark version: Leaderboard manifests checked 2026-09-30
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B; six public Arabic test sets under the Open Universal Arabic ASR Leaderboard protocol, equally weighted across test sets; lower WER/CER is better.
Context: TII does not publish per-set clip counts or the six individual Arabic test-set scores in this article.

Source TII Falcon-ASR article

Internal held-out Emirati evaluation WER (Show measurement, test conditions, and source)

Source value: 22.73
Score: 22.73
Metric: word error rate
Unit: %
Benchmark version: TII internal human-validated evaluation
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B on additional held-out Emirati/Gulf recordings; TII says transcripts were human-validated but gives neither sample count nor individual datasets.
Context: Internal vendor evaluation without a published dataset, sample count, or external replication.

Source TII Falcon-ASR article

Internal held-out Emirati evaluation CER (Show measurement, test conditions, and source)

Source value: 10.19
Score: 10.19
Metric: character error rate
Unit: %
Benchmark version: TII internal human-validated evaluation
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B on additional held-out Emirati/Gulf recordings; TII says transcripts were human-validated but gives neither sample count nor individual datasets.
Context: Internal vendor evaluation without a published dataset, sample count, or external replication.

Source TII Falcon-ASR article

English ASR WER: LibriSpeech clean (Show measurement, test conditions, and source)

Source value: 1.75
Score: 1.75
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set LibriSpeech clean. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: LibriSpeech other (Show measurement, test conditions, and source)

Source value: 4.21
Score: 4.21
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set LibriSpeech other. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: SPGISpeech (Show measurement, test conditions, and source)

Source value: 2.02
Score: 2.02
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set SPGISpeech. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: VoxPopuli (Show measurement, test conditions, and source)

Source value: 3.87
Score: 3.87
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set VoxPopuli. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: GigaSpeech (Show measurement, test conditions, and source)

Source value: 8.15
Score: 8.15
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set GigaSpeech. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: AMI (Show measurement, test conditions, and source)

Source value: 8.33
Score: 8.33
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set AMI. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

English ASR WER: Earnings-22 (Show measurement, test conditions, and source)

Source value: 11.86
Score: 11.86
Metric: word error rate
Unit: %
Benchmark version: Seven-set English evaluation described by TII
Category: audio
Direction: Lower is better
Source type: vendor-reported
Evaluator: TII Falcon-ASR article
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Falcon-ASR 1.6B with the same weights; public English test set Earnings-22. TII does not give this set's clip count.
Context: Vendor report; no run configuration, set size, or independent replication is published for this individual row.

Source TII Falcon-ASR article

Sources and data date

Type	Evidence and data date
Research status	Research date October 7, 2026. 1 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	arabic-sample-counts Original wording TII gives neither per-set clip counts nor the six individual WER/CER values for the public Arabic average.
Not published	emirati-evaluation-details Original wording The internal Emirati evaluation has no public manifest, sample count, per-set values, or independent replication.
Not found in the inspected sources	other-supported-language-benchmarks Original wording The TII source checked publishes benchmark results for Arabic and English, but none for French, Spanish, or Portuguese.
Additional source	TII Falcon model launch (retrieved October 7, 2026) · TII Falcon model launch · Editorial description reviewed October 7, 2026
Additional source	Falcon-ASR and Falcon-OCR-Arabic overview (retrieved October 7, 2026) · Falcon-ASR and Falcon-OCR-Arabic overview · Editorial description reviewed October 7, 2026
Additional source	Falcon-ASR public demo (retrieved October 7, 2026)
Additional source	TII Falcon-ASR article (retrieved October 7, 2026)