Datasets:
id stringlengths 10 13 | name stringlengths 4 64 | type stringclasses 6
values | layer int64 0 4 | source stringlengths 4 32 ⌀ | size stringlengths 4 33 ⌀ | hours float64 1.56 1k ⌀ | condition stringclasses 5
values | domain stringlengths 3 22 ⌀ | license stringlengths 3 34 ⌀ | link stringlengths 34 87 ⌀ | paper stringlengths 7 28 ⌀ | verified stringclasses 3
values | notes stringlengths 44 329 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
AZ-SPCORP-001 | WorldSpeech az_az | corpus | 2 | DISCO Lab, ETH Zurich | 305.4 h | 305.4 | broadcast | news - interviews | CC-BY-NC-4.0 | https://huggingface.co/datasets/disco-eth/WorldSpeech | arXiv:2605.09167 | documented | 92% of all documented-provenance AZ speech. Source is Voice of America, US federal public domain, so the same corpus can be lawfully re-derived for commercial use from source_url + timestamps in the card |
AZ-SPCORP-002 | LocalDoc/azerbaijani_asr | corpus | 2 | LocalDoc | 345,643 pairs | 328 | unstated | unstated | CC-BY-4.0 | https://huggingface.co/datasets/LocalDoc/azerbaijani_asr | null | claimed | Largest openly licensed AZ corpus. Source, speakers, collection method and transcript origin all undocumented - a licensing risk the CC-BY tag does not remove |
AZ-SPCORP-003 | FLEURS az_az | corpus | 1 | Google | 3,988 utts | 13.9 | read | FLoRes sentences | CC-BY-4.0 | https://huggingface.co/datasets/google/fleurs | arXiv:2205.12446 | measured | The de facto AZ benchmark. train 9.31h/2665, dev 1.35h/400, test 3.24h/923, measured directly. Two transcript defects found: U+0307 after i in ~10% of utts from locale-unaware lowercasing of capital I-dot, and unexpanded digits in ~23% |
AZ-SPCORP-004 | asmarhajizada/azerbaijani-audiobooks | corpus | 3 | individual | 7,513 utts | 27.1 | read | audiobooks, 7 titles | none declared | https://huggingface.co/datasets/asmarhajizada/azerbaijani-audiobooks | null | measured | No licence declared - do not use commercially |
AZ-SPCORP-005 | CMU Wilderness AZEBSA | corpus | 3 | CMU Festvox | 5,325 utts | 8.07 | read | New Testament | alignments only; audio restricted | https://github.com/festvox/datasets-CMU_Wilderness | 10.1109/ICASSP.2019.8683536 | documented | Until May 2026 this was the LARGEST human-transcribed open Azerbaijani corpus in existence |
AZ-SPCORP-006 | YODAS az000 | corpus | 3 | WavLab CMU | null | 3.85 | in-the-wild | YouTube | CC-BY-3.0 | https://huggingface.co/datasets/espnet/yodas | arXiv:2406.00899 | documented | Uploader-supplied captions, not necessarily human transcribed |
AZ-SPCORP-007 | Common Voice az | corpus | 1 | Mozilla | 1,036 clips / 48 speakers | 1.56 | read | crowd prompts | CC0-1.0 | https://commonvoice.mozilla.org/az | LREC 2020 | measured | v26.0 Jun 2026: 1.56h total, 0.65h VALIDATED. Turkish 130h, Russian 254h from the same effort. Accent field empty for every clip; variant mechanism never configured. Any published CV-az WER is computed on <=130 clips |
AZ-SPCORP-008 | VoxLingua107 az | corpus | 3 | TalTech | null | 58 | in-the-wild | YouTube | CC-BY-4.0 | https://bark.phon.ioc.ee/voxlingua107/ | arXiv:2011.12998 | documented | Language-ID labels only, NO transcripts - contributes nothing to ASR supervision |
AZ-SPCORP-009 | Emergency call-centre corpus | corpus | 4 | ATL Tech + ADA University + AzTU | null | 80 | telephone - spontaneous | 112 emergency calls | not released | null | 10.3390/sym13040634 | documented | THE ONLY documented Azerbaijani telephone and conversational corpus. 27h dialogue + 53h summaries, human transcribed, never released |
AZ-SPCORP-010 | yoyo-research-group/south-azerbaijani-asr | corpus | 3 | yoyo-research-group | 9,702 utts | 16.68 | read | 5 books | CC-BY-4.0 | https://huggingface.co/datasets/yoyo-research-group/south-azerbaijani-asr | null | claimed | The ONLY downloadable transcribed azb corpus. README is a single licence line. Speaker dirs end in -bot, leaving human-vs-synthetic unresolved |
AZ-SPCORP-011 | Kartal-Ol/AZB-ASR-Gold-Testset | benchmark | 1 | Kartalol | 3,079 recordings / 500 sentences | 17.24 | read | azb | CC-BY-NC-4.0 | https://huggingface.co/datasets/Kartal-Ol/AZB-ASR-Gold-Testset | INTERSPEECH 2026 | documented | First South Azerbaijani benchmark. Manifest published; AUDIO NOT RELEASED |
AZ-SPCORP-012 | AZ-SRDat | corpus | 4 | Institute of IT, ANAS | 86 speakers | null | read | speaker recognition | not released | null | Problems of IT 2013(1):67-73 | documented | Speaker recognition, not ASR. Never released |
AZ-SPCORP-013 | interneuronai/azspeech | corpus | 3 | Alas Development Center, Baku | 400,000+ files | 1,000 | unstated | web-scraped | preview Apache-2.0; full set gated | https://huggingface.co/datasets/interneuronai/azspeech | null | claimed | Vendor self-report. Only a 4-5k row preview is public. ~400h free to academics under agreement; commercial is paid |
AZ-ASR-001 | LocalDoc/azerbaijani-whisper-turbo | model | 1 | LocalDoc | whisper-large-v3-turbo FT | null | null | ASR | Apache-2.0 | https://huggingface.co/LocalDoc/azerbaijani-whisper-turbo | null | documented | BEST publicly reported AZ result: WER 13.17 / CER 3.45 on FLEURS az test (921 utts). Trained on the 328h LocalDoc corpus |
AZ-ASR-002 | nijatzeynalov/wav2vec2-large-mms-1b-azerbaijani-common_voice15.0 | model | 1 | Nijat Zeynalov | MMS-1B adapter FT | null | null | ASR | CC-BY-NC-4.0 | https://huggingface.co/nijatzeynalov/wav2vec2-large-mms-1b-azerbaijani-common_voice15.0 | null | documented | 394,384 downloads - roughly 98% of the entire Azerbaijani ASR ecosystem, against ~7,000 for all other Azerbaijani models combined. Three years old, non-commercial licence, WER 26.32 evaluated on ~29 clips |
AZ-ASR-003 | BuzzASR/azerbaijani | model | 1 | LEMN Lab | whisper-large-v3 FT | null | null | ASR | MIT | https://huggingface.co/BuzzASR/azerbaijani | Findings of EMNLP 2026 | documented | FLEURS 21.54 WER / 5.52 CER; CV25 12.10 / 3.05. One of 102 monolingual models |
AZ-ASR-004 | Kartal-Ol/ASR-AZB | model | 1 | Kartalol | 8 Whisper checkpoints | null | null | ASR azb | none set | https://huggingface.co/Kartal-Ol/ASR-AZB | INTERSPEECH 2026 | documented | The serious azb research line. Cross-lingual init ablation over fa/North-Az/tr/ar. Best GoldSet 70.0 WER; community set 22.0 with North-Azerbaijani init |
AZ-ASR-005 | BHOSAI/Pichilti-base-v1 | model | 2 | Baku Higher Oil School AI R&D | whisper-base, frozen encoder | null | null | ASR | CC-BY-SA-4.0 | https://huggingface.co/BHOSAI/Pichilti-base-v1 | pending | claimed | Only named Azerbaijani institution publishing ASR models. Trained on >500k unlabelled audios. No WER published |
AZ-ASR-006 | OpenAI Whisper (all sizes) | model | 1 | OpenAI | 39M-1.55B | null | null | multilingual ASR | MIT | https://huggingface.co/openai/whisper-large-v3 | arXiv:2212.04356 | documented | FLEURS az ladder: tiny 93.1, base 76.4, small 49.1, medium 33.1, large 28.7, large-v2 23.4. large-v3 19.7 appears only in a repo figure, in no paper. Absent from the Common Voice 9 table because CV9 az was 0.16h |
AZ-ASR-007 | facebook/mms-1b-all | model | 1 | Meta | 1B + adapters | null | null | 1,107-language ASR | CC-BY-NC-4.0 | https://huggingface.co/facebook/mms-1b-all | arXiv:2305.13516 | documented | Covers azj-script_latin, azj-script_cyrillic AND azb. FLEURS az 45.0 WER without an LM, 19.8 with (JMLR Table A1) - the LM is the decisive component. Best PEER-REVIEWED public result; the best KNOWN result, 13.17, is an unpublished model card. Non-commercial licence |
AZ-ASR-008 | Meta Omnilingual ASR | model | 1 | Meta | 1,600+ languages | null | null | multilingual ASR | Apache-2.0 | https://github.com/facebookresearch/omnilingual-asr | null | documented | Covers aze_Latn, aze_Cyrl AND aze_Arab. The ONLY permissively licensed model covering all three Azerbaijani scripts. Nov 2025 |
AZ-ASR-009 | facebook/seamless-m4t-v2-large | model | 2 | Meta | 2.3B | null | null | S2T | CC-BY-NC-4.0 | https://huggingface.co/facebook/seamless-m4t-v2-large | arXiv:2312.05187 | documented | Trained on 101h labelled azj ASR + 7,690h raw audio. Publishes NO per-language Azerbaijani WER, only 77-language aggregates |
AZ-SVC-001 | Soniox | service | 1 | commercial | $0.10 async / $0.12 realtime | null | null | cloud ASR | commercial | null | null | documented | Cheapest vendor that lists Azerbaijani. Documents that all supported languages work in both its real-time and async APIs. One of three vendors with documented streaming, alongside ElevenLabs and Gladia. Publishes no numeric Azerbaijani WER |
AZ-SVC-002 | SESTEK / Knovvu | service | 1 | commercial | unpriced | null | null | cloud ASR | commercial | null | null | documented | The ONLY vendor documenting a purpose-built Azerbaijani acoustic model, offered alongside Whisper and Dolphin backends. Also ships a bilingual Azerbaijani-Russian model. No AzerbaijaniStream model exists; streaming is a transport rather than a model. No rate published. Turkish company. Publishes no numeric Azerbaijani ... |
AZ-SVC-003 | Google STT v1 | service | 1 | commercial | $1.44 / $0.96 with logging | null | null | cloud ASR | commercial | null | null | documented | Most expensive and least capable AZ tier: no diarization, no model adaptation. Telephony models exist but az-AZ has only default and command_and_search. Publishes no numeric Azerbaijani WER |
AZ-SVC-004 | Google STT v2 / Chirp | service | 1 | commercial | $0.96 to $0.24 tiered | null | null | cloud ASR | commercial | null | null | documented | az on chirp_3 is Preview, not GA. Every az-AZ row is chirp family; none is telephony. Publishes no numeric Azerbaijani WER |
AZ-SVC-005 | Azure AI Speech | service | 1 | commercial | $1.00 realtime / $0.18 batch | null | null | cloud ASR | commercial | null | null | documented | Custom Speech restricted to plain-text adaptation for az. COUNTER-EXAMPLE to the 2026 narrowing trend: the newer MAI-Transcribe-2 covers Azerbaijani where MAI-Transcribe-1.5 did not, and az-AZ has gained fast-transcription support. Publishes no numeric Azerbaijani WER |
AZ-SVC-006 | Amazon Transcribe | service | 1 | commercial | $0.36 batch | null | null | cloud ASR | commercial | null | null | documented | Batch only. Marked NO for Call Analytics, its telephony product. Audio used for service improvement unless opted out at Organizations level. Publishes no numeric Azerbaijani WER |
AZ-SVC-007 | AssemblyAI | service | 1 | commercial | $0.15 async | null | null | cloud ASR | commercial | null | null | documented | Universal-2 only; EXCLUDED from Universal-3.5 Pro. Publishes a band, 10-25% WER. Marketing page promises accurate transcription across every dialect and lists North Azerbaijani, South Azerbaijani and Quba, while its own comparison table on the same page marks Azerbaijani as Universal-2 only. Publishes no numeric Azerba... |
AZ-SVC-008 | Deepgram | service | 1 | commercial | $0.288 | null | null | cloud ASR | commercial | null | null | documented | whisper-cloud only; absent from Nova-2, Nova-3 and the newer Flux. Documentation states explicitly that live streaming is unavailable on whisper-cloud. Publishes no numeric Azerbaijani WER |
AZ-SVC-009 | ElevenLabs Scribe v2 | service | 1 | commercial | $0.22 batch / $0.39 realtime | null | null | cloud ASR | commercial | null | null | documented | The ONLY vendor publishing a numeric Azerbaijani figure: 9.4% WER on FLEURS for Scribe v1, on its product page, alongside competitor figures. Its developer documentation publishes only bands and places Azerbaijani in 5-10%, disagreeing with AssemblyAI by a full tier. Publishes no numeric Azerbaijani WER |
AZ-SVC-010 | Gladia | service | 1 | commercial | $0.61 async PAYG / $0.75 realtime | null | null | cloud ASR | commercial | null | null | documented | Solaria-1 supports Azerbaijani in both async and live. Solaria-3 does not, but it is a five-language European model rather than a flagship that dropped the language. PAYG tier lists no training opt-out. Publishes no numeric Azerbaijani WER |
AZ-SVC-011 | Rev.ai | service | 1 | commercial | ~$0.30 | null | null | cloud ASR | commercial | null | null | documented | Async only. Rev states it trains its own ASR models on customer audio. Publishes no numeric Azerbaijani WER |
AZ-SVC-012 | OpenAI whisper-1 | service | 1 | commercial | $0.36 | null | null | cloud ASR | commercial | null | null | documented | Stock Whisper behind an API. Publishes no numeric Azerbaijani WER |
AZ-GAP-001 | Vosk | gap | 0 | null | null | null | null | null | null | null | null | documented | NO Azerbaijani model. Verified three ways: the model list has Turkish, Kazakh, Uzbek, Kyrgyz, Tajik and Farsi but not Azerbaijani; a GitHub issue search returns zero results, so nobody has even asked; no az lexicon ships |
AZ-GAP-002 | NVIDIA NeMo | gap | 0 | null | null | null | null | null | null | null | null | documented | No az checkpoint anywhere. Enumerating the org's speech-recognition checkpoints finds Kazakh, Uzbek, Georgian and Armenian but no Azerbaijani, and a catalogue search returns nothing. Canary, Parakeet and Granary are 25 European languages |
AZ-GAP-003 | Kaldi | gap | 0 | null | null | null | null | null | null | null | null | documented | No Azerbaijani recipe among 104 egs directories |
AZ-GAP-004 | Coqui STT | gap | 0 | null | null | null | null | null | null | null | null | documented | 54 languages across 92 releases, no az. Project archived |
AZ-GAP-005 | Speechmatics | gap | 0 | null | null | null | null | null | null | null | null | documented | Language list includes Bashkir and Turkish, not Azerbaijani |
AZ-GAP-006 | Yandex SpeechKit | gap | 0 | null | null | null | null | null | null | null | null | documented | Ships Turkish, Kazakh and Uzbek and explicitly skips Azerbaijani |
AZ-GAP-007 | Alibaba / Qwen ASR | gap | 0 | null | null | null | null | null | null | null | null | documented | Absent across all four surfaces. Intelligent Speech Interaction ships Kazakh, not Azerbaijani |
AZ-GAP-008 | xAI Grok STT 1.0 | gap | 0 | null | null | null | null | null | null | null | null | documented | Explicit 25-language table with tr and fa but not az. MEASURED producing Turkish orthography: LSGR 0.000 on telephone speech |
AZ-TOOL-001 | LSGR | tool | 1 | ALMAZ | null | null | null | evaluation metric | CC-BY-4.0 | https://huggingface.co/datasets/almaz-nlp/lsgr | 10.5281/zenodo.22742916 | measured | Language-Specific Grapheme Recall. Detects a system writing a neighbouring language's orthography. Its telephone evaluation appears to be the only Azerbaijani telephone-band ASR evaluation in existence |
AZ-TOOL-002 | MorAz | tool | 2 | Ozenc, Ehsani, Solak | null | null | null | morphological analyser | open source | null | 10.18653/v1/D18-2005 | documented | The only Azerbaijani morphological analyser with a paper. Relevant to ASR because agglutination drives the out-of-vocabulary problem |
The ALMAZ ASR Roster
Archived at Zenodo: 10.5281/zenodo.22761871 (concept DOI, always resolves to the latest version).
A curated catalog of Azerbaijani speech-to-text artifacts: corpora, models, services, benchmarks and tools. Companion to the ALMAZ Resource Roster, which does the same for text.
Schema matches the text roster so the two join, plus three columns speech needs
and text does not: hours, condition, and verified.
The verified column
The standard way a survey goes wrong is laundering vendor claims into apparent facts by citing them. Every row says which it is:
| value | meaning |
|---|---|
measured |
we ran it or counted it ourselves |
documented |
stated in the owner's own documentation, which we read |
claimed |
asserted without evidence we could check |
What the roster shows
- One vendor in thirteen publishes a numeric Azerbaijani accuracy figure, and on a marketing page rather than in its documentation. Two others publish coarse bands that disagree by a full tier.
- Exactly one vendor documents a purpose-built Azerbaijani acoustic model, and it is Turkish. Everything else is Whisper or MMS under a brand.
- There is no public spontaneous or telephone-band Azerbaijani speech. The only telephone corpus that exists, 80 h of emergency calls, was never released.
- Every natively-8 kHz open stack has no Azerbaijani model — Vosk, Kaldi, NeMo, Coqui — while every stack that has Azerbaijani is 16 kHz and offline.
- One three-year-old model is 98% of the ecosystem's downloads, evaluated on about 29 clips.
- Common Voice Azerbaijani has 39 minutes of validated audio, against 130 hours for Turkish from the same effort.
- South Azerbaijani has roughly 17 h against 330–690 h for the north, and no commercial support anywhere.
Provenance
Vendor rows were re-verified against each company's own live documentation on 15 September 2026. Seven verified absences were each confirmed by at least two independent probes. Where a marketing page and a technical document disagree, the roster records the technical document and notes the conflict.
Caveats that matter
Hours do not sum. The two largest open corpora share splits with FLEURS and with each other, so they are partly aggregates rather than independent resources. Download counts, prices and language lists were retrieved on 15 September 2026 and move constantly.
Citation
Part of the ALMAZ programme. See also LSGR, the orthographic conformance metric used to separate claimed from demonstrated language support (doi:10.5281/zenodo.22742916).
- Downloads last month
- 52