Catalogue · বাংলা
Ready-made Bengali speech datasets
Corpora already recorded in Bengali and available under licence, plus what to check before buying existing data instead of commissioning a build.
- Language
- Bengali (bn-IN)
- Licence
- Perpetual, commercial, model-training permitted
- Availability
- Subject to consent scope; confirmed on enquiry
- Lead time
- Days, not weeks
Available corpus types
| Corpus | Typical band | Annotation | Best for |
|---|---|---|---|
| Bengali scripted speech | 100-500 hours | Verbatim + timestamps | Prompt-read speech from a phonetically balanced script, the baseline corpus type for ASR and TTS training. |
| Bengali spontaneous speech | 100-500 hours | Verbatim + timestamps | Unscripted monologue on prompted topics, which carries the disfluencies and prosody that scripted data never produces. |
| Bengali conversational speech | 100-500 hours | Verbatim + speaker turns + timestamps | Two-party conversation recorded on separate channels, with overlap and turn-taking preserved. |
| Bengali telephony speech | 50-300 hours | Verbatim + timestamps | Narrowband call audio captured over a telephony path, matching what a deployed contact-centre model actually receives. |
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.
Ready-made versus commissioned
| Ready-made | Commissioned build | |
|---|---|---|
| Lead time | Days | 3-16 weeks |
| Cost per hour | Lower | Higher |
| Cohort control | Fixed by the original spec | Yours |
| Dialect quotas | As recorded | Designed |
| Exclusivity | Non-exclusive | Exclusive if you want it |
| Best for | Baselines, prototypes, augmentation | Production models |
What to check before licensing existing data
- 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 West Bengal, Tripura, Assam (Barak Valley)
- Recording conditions: studio-only data underperforms badly on noisy deployments
- Transcription convention: Three sibilant characters chosen inconsistently for the same sound will otherwise surface as tokenisation noise
- Overlap: check whether the corpus already sits in the public sets you trained on
Where the public corpora fall short
Indian Bengali is under-collected relative to Bangladeshi Bengali, and North Bengal and Tripura varieties are almost entirely missing.
This is usually the reason buyers move from licensing to commissioning: the free and cheap material covers the easy half of the language.
What a usable Bengali corpus has to cover
- Dialect spread across Kolkata standard (Rarhi), Sylheti-influenced, Rangpuri / North Bengal, Medinipuri rather than a single prestige variety
- The phonetic contrasts that Bengali models actually get wrong: Inherent vowel is realised as /ɔ/ or /o/, which breaks G2P rules copied from Devanagari-based systems
- Code-mixed speech transcribed rather than discarded — Kolkata professional speech mixes English heavily; rural West Bengal much less. A single 'Bengali' dataset without register tags conflates two very different acoustic and lexical distributions.
- Recording geography covering Kolkata, Siliguri, Durgapur, Agartala, since dialect follows location
- Utterance types beyond read prompts: Cholit-bhasha spontaneous conversation, Read prompts covering cluster simplification, Regional dialect sets from North Bengal and Tripura
Tag every speaker as Indian Bengali and record district of origin; mixing in Bangladeshi speech without tags is a common and costly dataset defect.
Frequently asked
Can I buy an existing Bengali dataset today?
Where a corpus exists within an appropriate consent and licence scope, yes — usually within days. We confirm availability against your intended use before quoting.
Is the licence exclusive?
Ready-made corpora are licensed non-exclusively. Exclusivity is available on commissioned builds.
Can we augment a ready-made corpus with new recordings?
That is the most common pattern: licence the base, then commission the dialects, conditions or domain vocabulary it lacks.
Do you provide a data sheet?
Yes — speaker counts, demographic distribution, condition mix, annotation convention and known limitations, before you commit.
Related pages
Check Bengali availability
Tell us the hours, corpus type and intended use. We confirm what is licensable now and what needs recording.