Ready-made Malayalam speech datasets

Availability is confirmed per enquiry. Some material can only be licensed within the consent scope the speakers originally agreed to, and we will not stretch that to close a sale.

• Consent scope: does it permit commercial model training, and by a third party?
• Dialect distribution: a corpus that is 90% one urban variety will not generalise across Kerala, Lakshadweep, Puducherry (Mahe)
• Recording conditions: studio-only data underperforms badly on noisy deployments
• Transcription convention: Old vs new script (chillu characters, Unicode normalisation) mixed within a dataset will otherwise surface as tokenisation noise
• Overlap: check whether the corpus already sits in the public sets you trained on

Central Kerala news-reading dominates public data. Malabar and southern varieties, and fast conversational speech generally, are missing.

This is usually the reason buyers move from licensing to commissioning: the free and cheap material covers the easy half of the language.

• Dialect spread across Thiruvananthapuram, Kochi (central), Malabar / Kozhikode, Thrissur rather than a single prestige variety
• The phonetic contrasts that Malayalam models actually get wrong: One of the most consonant-dense Indian languages; long geminates and clusters raise word error rates sharply
• Code-mixed speech transcribed rather than discarded — Manglish is standard in urban and professional speech, with heavy English noun and verb insertion.
• Recording geography covering Kochi, Thiruvananthapuram, Kozhikode, Thrissur, since dialect follows location
• Utterance types beyond read prompts: Fast spontaneous conversation with verbatim transcription, Read prompts covering geminates and clusters, Mappila Malayalam dialect sets from Kozhikode and Malappuram

Budget higher transcription effort per audio hour for Malayalam than for Hindi; speech rate and morphology make it slower to annotate.