TembraAIPlatform/fleurs-armenian-corrections

Errors in the Armenian FLEURS test transcripts

FLEURS (google/fleurs, config hy_am) is one of the main public test sets for Armenian speech recognition. Its original transcripts are the sentences the readers were given to read, not a record of what they said. For ArmBench-ASR, Metric AI Lab corrected the test transcripts by hand and published corrected transcripts for 314 test recordings (Metric-AI/fleurs-corrections). Our benchmark of speech recognition for low-resource languages scores Armenian FLEURS against these ArmBench transcripts. On top of them, we went through every place where a leading system — Tembra, ElevenLabs Scribe v2, Google Chirp 3 or NVIDIA stt_hy — disagrees with the transcript and checked the differences in wording against the recording. This page shows how much the transcripts change the measured error rate and lists the 13 further corrections we found.

Summary

• Digits and Latin script. Many original transcripts have digits (such as «6-6» or «29¾») or Latin letters: abbreviations and foreign words such as «pH», «SANParks» or «Wild Card», Latin words such as «civis», and even HTML markup (<sup> </sup>). A system can write these places with digits and Latin letters or in Armenian words, and the recording does not tell which is right, so on these recordings WER mostly measures how a system writes numbers and foreign words. We do not score recordings whose reference (ArmBench) transcript has digits or Latin letters, for any system.
• ArmBench corrections. 301 of the 314 ArmBench corrections change words, not only punctuation or letter case.
• Our further corrections. In 12 recordings the reader says something other than the ArmBench transcript: 13 places. Two of them are errors in the transcript itself (a typo and a wrong article); in the others the reader says a different word or form than the transcript, or skips a word.
• Effect on WER. On the scored test set (utterances without digits or Latin-script words) the ArmBench corrections lower WER by 2.3–4.2 points, depending on the system; our 13 corrections lower it further for every system, by 0.03–0.27 points. Tembra has the lowest WER with each of the three versions of the transcripts — down to 2.18 % WER and 0.55 % CER. Of the 407 candidate places, in 270 the recording matches the transcript, 13 were corrected, and 124 were deliberately left as they are under the strict rules below: differences in spelling, pronunciation, names or numbers, or places that did not pass every check.

Effect on WER and CER

WER and character error rate (CER), %, against the original FLEURS transcripts, the ArmBench transcripts, and the ArmBench transcripts with our 13 corrections; 95 % bootstrap confidence intervals (2,000 resamples over FLEURS sentences) in parentheses. The same transcripts are used for every system.

system	WER, original FLEURS	WER, ArmBench	WER, ArmBench + our corrections	CER, original FLEURS	CER, ArmBench	CER, ArmBench + our corrections
Tembra	6.63 (5.3–8.3)	2.45 (1.7–3.3)	2.18 (1.5–3.0)	2.04	0.61	0.55
ElevenLabs Scribe v2	7.99 (6.7–9.3)	3.90 (3.0–4.9)	3.71 (2.9–4.7)	2.41	1.12	1.07
Google Chirp 3	7.82 (6.6–9.0)	4.25 (3.3–5.4)	4.09 (3.1–5.2)	2.40	1.47	1.42
NVIDIA stt_hy FastConformer	10.69 (9.1–12.4)	6.64 (5.5–7.9)	6.54 (5.3–7.7)	3.24	1.70	1.66
OpenAI Whisper large-v3	40.96 (38.7–43.2)	38.65 (36.5–40.8)	38.62 (36.4–40.9)	9.14	8.03	8.01

Our 13 corrections

Texts are normalized as in scoring (ArmBench-ASR normalization: lowercase, no punctuation, number words written as digits). Every entry can be checked by listening to the FLEURS recording.

FLEURS sentence id	audio file	in context: [ArmBench transcript → reader says]	kind
1990	12163641903107321106.wav	… ուսանողները հաճախ իրենց բլոգի [փորձառությունն → փորձառությունը] սկսում են անփույթ քրեական …	wrong article in the transcript
1874	13082929885147911800.wav	… նույնիսկ մյուս աստղերը պտտվում [էին → են] երկրի շուրջ …	the reader says another word or form
1742	14078831229922756456.wav	… անունն է տալիս հոնկոնգի [տարածքի → տարածքին] եւ բազմաթիվ զբոսաշրջիկների ուշադրության …	the reader says another word or form
1896	15235462427526863845.wav	… երազանքներին հասնել անվտանգ դրական [եւ → —] աջակցող միջավայրում …	the reader skips the word
1899	15466465960990209111.wav	… խորապես հետաքրքրված էր այս [աշխարհիկ → աշխարհի] կյանքից դուրս գտնվող հարցերով …	the reader says another word or form
1971	15992141011392489499.wav	… քամիները կարկուտը չափից շատ [տեղումների → տեղումները] եւ անտառային հրդեհները դաժան …	the reader says another word or form
1831	17356326692173544167.wav	… է հիմնականում հաճարենիներով փշատերեւ [ծառերով ու → ծառերով եւ] եղեւնիներով եւ ալպյան ու …	the reader says another word or form
1831	17356326692173544167.wav	… ու եղեւնիներով եւ ալպյան [ու → եւ] միջերկրածովյան բուսականության խառնուրդ է …	the reader says another word or form
1952	2295828259369518648.wav	… բարձր լայնություններում կամ լեռնանցքների [վրա → վրայով] պետք է դիտարկի ձյունը …	the reader says another word or form
1714	3543763767166930323.wav	… քանի որ երկրի մթնոլորտն [աղավաղում → աղավաղվում] է մեր լույսի եւ …	the reader says another word or form
1961	4074491694262354091.wav	… դիտարկումն է ովքեր անցնելով [արհամարանքի → արհամարհանքի] դժբախտության եւ չափահասների կողմից …	typo in the transcript
1730	6293805952214621769.wav	… երկայնքով չնայած արեւելք եւ [ջաքար բումթանգ գնացողները → ջաքար բումթանգ գնացող] մեկնում են 6 անց …	the reader says another word or form
1944	7995091563962156953.wav	… բազմաթիվ տարատեսակ խանութներով լողալն [ապահով → անապահով] է …	the reader says another word or form

How the recordings were checked

1. Recordings. The test recordings scored on our benchmark page: utterances whose reference (ArmBench) transcripts have no digits and no Latin-script words.
2. Candidate places. The transcripts and the output of several speech recognition systems, including Tembra, ElevenLabs Scribe v2, Google Chirp 3 and NVIDIA stt_hy, were normalized and aligned word by word. Every place where at least one of them differs from the transcript is a candidate: 407 places.
3. Places that sound the same. 59 candidates differ only in spelling or in the spelling of a proper name and sound the same; they were left as they are.
4. Verification against the recording. Each of the other 348 places was checked against the recording with the competing variants in random order and without the information which system wrote which.
5. When a place was corrected. Only when all of these hold:
   • the recording clearly says a variant that differs from the transcript;
   • it is another word or word form: a one-letter difference that is only pronunciation stays as it is, unless the transcript word is a typo (a form that practically never occurs in a large Armenian text corpus, while the form in the recording does);
   • proper names and numbers stay as they are;
   • at least two of the four systems named above write the corrected variant, so no single system, ours included, can change a transcript on its own.
6. Outcome. 13 corrections; 270 places where the recording matches the transcript; 65 kept unchanged because not all of the rules above hold.

Files

• data/corrections.jsonl — our 13 corrections: fleurs_sentence_id, audio_file, fleurs_original_transcript, armbench_transcript, armbench_transcript_normalized, in_transcript (the place in the ArmBench transcript), in_original_transcript (the same place in the original transcript), said (what the reader says; empty where nothing is said), kind, vs_original (same, original_matches or all_differ), corrected_transcript.

fleurs_original_transcript is the transcript as published in google/fleurs; armbench_transcript is the transcript with the ArmBench-ASR corrections; corrected_transcript adds our corrections and, like armbench_transcript_normalized, is normalized as in scoring. To apply the corrections, replace the transcript of each listed recording with corrected_transcript and normalize system output the same way.

Credits, license and citation

The transcripts are derived from FLEURS by Google (google/fleurs, CC BY 4.0) and from the ArmBench-ASR corrections by Metric AI Lab (Metric-AI/fleurs-corrections, CC BY 4.0); this data is released under CC BY 4.0. Please cite FLEURS, ArmBench-ASR and this dataset (Tembra AI, 2026, https://huggingface.co/datasets/TembraAIPlatform/fleurs-armenian-corrections):

@article{conneau2022fleurs,
  title={FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author={Conneau, Alexis and Ma, Min and Khanuja, Simran and Zhang, Yu and Axelrod, Vera and Dalmia, Siddharth and Riesa, Jason and Rivera, Clara and Bapna, Ankur},
  journal={arXiv preprint arXiv:2205.12446},
  year={2022}
}

ArmBench-ASR: Alexander Shahramanyan, Mariam Avetisyan, Hrant Davtyan (Metric AI Lab), 2026 — https://huggingface.co/blog/Metric-AI/armbench-asr

Contact

Tembra AI — https://tembra.ai