| |
| """Build the ALMAZ ASR Roster. |
| |
| Schema mirrors almaz-nlp/almaz-roster so the text and speech rosters join on |
| the same columns, plus three that speech needs and text does not: hours, |
| condition, and verified. |
| |
| `verified` is the integrity column. A survey's standard failure is laundering |
| vendor claims into apparent facts by citing them, so every row says whether we |
| ran it (measured), read it in the owner's own documentation (documented), or |
| are repeating an unevidenced assertion (claimed). |
| """ |
| import csv |
| from pathlib import Path |
|
|
| COLS = ["id","name","type","layer","source","size","hours","condition", |
| "domain","license","link","paper","verified","notes"] |
|
|
| R = [] |
| def add(**kw): |
| R.append({c: kw.get(c, "") for c in COLS}) |
|
|
| |
| add(id="AZ-SPCORP-001", name="WorldSpeech az_az", type="corpus", layer=2, source="DISCO Lab, ETH Zurich", |
| size="305.4 h", hours=305.4, condition="broadcast", domain="news - interviews", license="CC-BY-NC-4.0", |
| link="https://huggingface.co/datasets/disco-eth/WorldSpeech", paper="arXiv:2605.09167", verified="documented", |
| notes="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") |
| add(id="AZ-SPCORP-002", name="LocalDoc/azerbaijani_asr", type="corpus", layer=2, source="LocalDoc", |
| size="345,643 pairs", hours=328, condition="unstated", domain="unstated", license="CC-BY-4.0", |
| link="https://huggingface.co/datasets/LocalDoc/azerbaijani_asr", paper="", verified="claimed", |
| notes="Largest openly licensed AZ corpus. Source, speakers, collection method and transcript origin all undocumented - a licensing risk the CC-BY tag does not remove") |
| add(id="AZ-SPCORP-003", name="FLEURS az_az", type="corpus", layer=1, source="Google", |
| size="3,988 utts", hours=13.9, condition="read", domain="FLoRes sentences", license="CC-BY-4.0", |
| link="https://huggingface.co/datasets/google/fleurs", paper="arXiv:2205.12446", verified="measured", |
| notes="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%") |
| add(id="AZ-SPCORP-004", name="asmarhajizada/azerbaijani-audiobooks", type="corpus", layer=3, source="individual", |
| size="7,513 utts", hours=27.1, condition="read", domain="audiobooks, 7 titles", license="none declared", |
| link="https://huggingface.co/datasets/asmarhajizada/azerbaijani-audiobooks", paper="", verified="measured", |
| notes="No licence declared - do not use commercially") |
| add(id="AZ-SPCORP-005", name="CMU Wilderness AZEBSA", type="corpus", layer=3, source="CMU Festvox", |
| size="5,325 utts", hours=8.07, condition="read", domain="New Testament", license="alignments only; audio restricted", |
| link="https://github.com/festvox/datasets-CMU_Wilderness", paper="10.1109/ICASSP.2019.8683536", verified="documented", |
| notes="Until May 2026 this was the LARGEST human-transcribed open Azerbaijani corpus in existence") |
| add(id="AZ-SPCORP-006", name="YODAS az000", type="corpus", layer=3, source="WavLab CMU", |
| size="", hours=3.85, condition="in-the-wild", domain="YouTube", license="CC-BY-3.0", |
| link="https://huggingface.co/datasets/espnet/yodas", paper="arXiv:2406.00899", verified="documented", |
| notes="Uploader-supplied captions, not necessarily human transcribed") |
| add(id="AZ-SPCORP-007", name="Common Voice az", type="corpus", layer=1, source="Mozilla", |
| size="1,036 clips / 48 speakers", hours=1.56, condition="read", domain="crowd prompts", license="CC0-1.0", |
| link="https://commonvoice.mozilla.org/az", paper="LREC 2020", verified="measured", |
| notes="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") |
| add(id="AZ-SPCORP-008", name="VoxLingua107 az", type="corpus", layer=3, source="TalTech", |
| size="", hours=58, condition="in-the-wild", domain="YouTube", license="CC-BY-4.0", |
| link="https://bark.phon.ioc.ee/voxlingua107/", paper="arXiv:2011.12998", verified="documented", |
| notes="Language-ID labels only, NO transcripts - contributes nothing to ASR supervision") |
| add(id="AZ-SPCORP-009", name="Emergency call-centre corpus", type="corpus", layer=4, source="ATL Tech + ADA University + AzTU", |
| size="", hours=80, condition="telephone - spontaneous", domain="112 emergency calls", license="not released", |
| link="", paper="10.3390/sym13040634", verified="documented", |
| notes="THE ONLY documented Azerbaijani telephone and conversational corpus. 27h dialogue + 53h summaries, human transcribed, never released") |
| add(id="AZ-SPCORP-010", name="yoyo-research-group/south-azerbaijani-asr", type="corpus", layer=3, source="yoyo-research-group", |
| size="9,702 utts", hours=16.68, condition="read", domain="5 books", license="CC-BY-4.0", |
| link="https://huggingface.co/datasets/yoyo-research-group/south-azerbaijani-asr", paper="", verified="claimed", |
| notes="The ONLY downloadable transcribed azb corpus. README is a single licence line. Speaker dirs end in -bot, leaving human-vs-synthetic unresolved") |
| add(id="AZ-SPCORP-011", name="Kartal-Ol/AZB-ASR-Gold-Testset", type="benchmark", layer=1, source="Kartalol", |
| size="3,079 recordings / 500 sentences", hours=17.24, condition="read", domain="azb", license="CC-BY-NC-4.0", |
| link="https://huggingface.co/datasets/Kartal-Ol/AZB-ASR-Gold-Testset", paper="INTERSPEECH 2026", verified="documented", |
| notes="First South Azerbaijani benchmark. Manifest published; AUDIO NOT RELEASED") |
| add(id="AZ-SPCORP-012", name="AZ-SRDat", type="corpus", layer=4, source="Institute of IT, ANAS", |
| size="86 speakers", hours="", condition="read", domain="speaker recognition", license="not released", |
| link="", paper="Problems of IT 2013(1):67-73", verified="documented", |
| notes="Speaker recognition, not ASR. Never released") |
| add(id="AZ-SPCORP-013", name="interneuronai/azspeech", type="corpus", layer=3, source="Alas Development Center, Baku", |
| size="400,000+ files", hours=1000, condition="unstated", domain="web-scraped", license="preview Apache-2.0; full set gated", |
| link="https://huggingface.co/datasets/interneuronai/azspeech", paper="", verified="claimed", |
| notes="Vendor self-report. Only a 4-5k row preview is public. ~400h free to academics under agreement; commercial is paid") |
|
|
| |
| add(id="AZ-ASR-001", name="LocalDoc/azerbaijani-whisper-turbo", type="model", layer=1, source="LocalDoc", |
| size="whisper-large-v3-turbo FT", hours="", condition="", domain="ASR", license="Apache-2.0", |
| link="https://huggingface.co/LocalDoc/azerbaijani-whisper-turbo", paper="", verified="documented", |
| notes="BEST publicly reported AZ result: WER 13.17 / CER 3.45 on FLEURS az test (921 utts). Trained on the 328h LocalDoc corpus") |
| add(id="AZ-ASR-002", name="nijatzeynalov/wav2vec2-large-mms-1b-azerbaijani-common_voice15.0", type="model", layer=1, |
| source="Nijat Zeynalov", size="MMS-1B adapter FT", hours="", condition="", domain="ASR", license="CC-BY-NC-4.0", |
| link="https://huggingface.co/nijatzeynalov/wav2vec2-large-mms-1b-azerbaijani-common_voice15.0", paper="", verified="documented", |
| notes="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") |
| add(id="AZ-ASR-003", name="BuzzASR/azerbaijani", type="model", layer=1, source="LEMN Lab", |
| size="whisper-large-v3 FT", hours="", condition="", domain="ASR", license="MIT", |
| link="https://huggingface.co/BuzzASR/azerbaijani", paper="Findings of EMNLP 2026", verified="documented", |
| notes="FLEURS 21.54 WER / 5.52 CER; CV25 12.10 / 3.05. One of 102 monolingual models") |
| add(id="AZ-ASR-004", name="Kartal-Ol/ASR-AZB", type="model", layer=1, source="Kartalol", |
| size="8 Whisper checkpoints", hours="", condition="", domain="ASR azb", license="none set", |
| link="https://huggingface.co/Kartal-Ol/ASR-AZB", paper="INTERSPEECH 2026", verified="documented", |
| notes="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") |
| add(id="AZ-ASR-005", name="BHOSAI/Pichilti-base-v1", type="model", layer=2, source="Baku Higher Oil School AI R&D", |
| size="whisper-base, frozen encoder", hours="", condition="", domain="ASR", license="CC-BY-SA-4.0", |
| link="https://huggingface.co/BHOSAI/Pichilti-base-v1", paper="pending", verified="claimed", |
| notes="Only named Azerbaijani institution publishing ASR models. Trained on >500k unlabelled audios. No WER published") |
| add(id="AZ-ASR-006", name="OpenAI Whisper (all sizes)", type="model", layer=1, source="OpenAI", |
| size="39M-1.55B", hours="", condition="", domain="multilingual ASR", license="MIT", |
| link="https://huggingface.co/openai/whisper-large-v3", paper="arXiv:2212.04356", verified="documented", |
| notes="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") |
| add(id="AZ-ASR-007", name="facebook/mms-1b-all", type="model", layer=1, source="Meta", |
| size="1B + adapters", hours="", condition="", domain="1,107-language ASR", license="CC-BY-NC-4.0", |
| link="https://huggingface.co/facebook/mms-1b-all", paper="arXiv:2305.13516", verified="documented", |
| notes="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") |
| add(id="AZ-ASR-008", name="Meta Omnilingual ASR", type="model", layer=1, source="Meta", |
| size="1,600+ languages", hours="", condition="", domain="multilingual ASR", license="Apache-2.0", |
| link="https://github.com/facebookresearch/omnilingual-asr", paper="", verified="documented", |
| notes="Covers aze_Latn, aze_Cyrl AND aze_Arab. The ONLY permissively licensed model covering all three Azerbaijani scripts. Nov 2025") |
| add(id="AZ-ASR-009", name="facebook/seamless-m4t-v2-large", type="model", layer=2, source="Meta", |
| size="2.3B", hours="", condition="", domain="S2T", license="CC-BY-NC-4.0", |
| link="https://huggingface.co/facebook/seamless-m4t-v2-large", paper="arXiv:2312.05187", verified="documented", |
| notes="Trained on 101h labelled azj ASR + 7,690h raw audio. Publishes NO per-language Azerbaijani WER, only 77-language aggregates") |
|
|
| |
| for i,(nm,price,stream,note) in enumerate([ |
| ("Soniox","$0.10 async / $0.12 realtime","yes", |
| "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"), |
| ("SESTEK / Knovvu","unpriced","unknown", |
| "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"), |
| ("Google STT v1","$1.44 / $0.96 with logging","probable", |
| "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"), |
| ("Google STT v2 / Chirp","$0.96 to $0.24 tiered","no on chirp_2", |
| "az on chirp_3 is Preview, not GA. Every az-AZ row is chirp family; none is telephony"), |
| ("Azure AI Speech","$1.00 realtime / $0.18 batch","likely", |
| "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"), |
| ("Amazon Transcribe","$0.36 batch","no", |
| "Batch only. Marked NO for Call Analytics, its telephony product. Audio used for service improvement unless opted out at Organizations level"), |
| ("AssemblyAI","$0.15 async","no", |
| "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"), |
| ("Deepgram","$0.288","no", |
| "whisper-cloud only; absent from Nova-2, Nova-3 and the newer Flux. Documentation states explicitly that live streaming is unavailable on whisper-cloud"), |
| ("ElevenLabs Scribe v2","$0.22 batch / $0.39 realtime","yes", |
| "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"), |
| ("Gladia","$0.61 async PAYG / $0.75 realtime","yes", |
| "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"), |
| ("Rev.ai","~$0.30","no","Async only. Rev states it trains its own ASR models on customer audio"), |
| ("OpenAI whisper-1","$0.36","no","Stock Whisper behind an API"), |
| ]): |
| add(id=f"AZ-SVC-{i+1:03d}", name=nm, type="service", layer=1, source="commercial", |
| size=price, hours="", condition="", domain="cloud ASR", license="commercial", |
| link="", paper="", verified="documented", |
| notes=note + ". Publishes no numeric Azerbaijani WER") |
|
|
| |
| for i,(nm,note) in enumerate([ |
| ("Vosk","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"), |
| ("NVIDIA NeMo","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"), |
| ("Kaldi","No Azerbaijani recipe among 104 egs directories"), |
| ("Coqui STT","54 languages across 92 releases, no az. Project archived"), |
| ("Speechmatics","Language list includes Bashkir and Turkish, not Azerbaijani"), |
| ("Yandex SpeechKit","Ships Turkish, Kazakh and Uzbek and explicitly skips Azerbaijani"), |
| ("Alibaba / Qwen ASR","Absent across all four surfaces. Intelligent Speech Interaction ships Kazakh, not Azerbaijani"), |
| ("xAI Grok STT 1.0","Explicit 25-language table with tr and fa but not az. MEASURED producing Turkish orthography: LSGR 0.000 on telephone speech"), |
| ]): |
| add(id=f"AZ-GAP-{i+1:03d}", name=nm, type="gap", layer=0, source="", |
| size="", hours="", condition="", domain="", license="", |
| link="", paper="", verified="documented", notes=note) |
|
|
| |
| add(id="AZ-TOOL-001", name="LSGR", type="tool", layer=1, source="ALMAZ", |
| size="", hours="", condition="", domain="evaluation metric", license="CC-BY-4.0", |
| link="https://huggingface.co/datasets/almaz-nlp/lsgr", paper="10.5281/zenodo.22742916", verified="measured", |
| notes="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") |
| add(id="AZ-TOOL-002", name="MorAz", type="tool", layer=2, source="Ozenc, Ehsani, Solak", |
| size="", hours="", condition="", domain="morphological analyser", license="open source", |
| link="", paper="10.18653/v1/D18-2005", verified="documented", |
| notes="The only Azerbaijani morphological analyser with a paper. Relevant to ASR because agglutination drives the out-of-vocabulary problem") |
|
|
| out = Path(__file__).parent / "almaz_asr_roster.csv" |
| with out.open("w", newline="", encoding="utf-8") as fh: |
| w = csv.DictWriter(fh, fieldnames=COLS) |
| w.writeheader(); w.writerows(R) |
|
|
| from collections import Counter |
| print(f"wrote {out.name}: {len(R)} entries") |
| print(" vendor rows re-verified against live documentation on 15 September 2026") |
| print(" by type:", dict(Counter(r['type'] for r in R))) |
| print(" by verification:", dict(Counter(r['verified'] for r in R if r['verified']))) |
| h = [float(r['hours']) for r in R if r['hours'] and r['type']=='corpus'] |
| print(f" corpora with hours: {len(h)}, total {sum(h):.1f} h") |
|
|
