Models and languages

The recommended transcription model for each of 10 Indian languages, plus a Hinglish compatibility code.

Pass a model name to model in the form field for non-streaming, in the config frame for streaming. There is no auto-detection: the model you name determines the language.

Available models

Language	Code	Model	Notes
Bengali	`bn`	`bn-banking-v2-8khz`	Bilingual with English (code-switching)
English	`en`	`en-banking-v2-8khz`	English-only (dedicated model)
Gujarati	`gu`	`gu-banking-v2-8khz`	Monolingual
Hindi	`hi`	`hi-banking-v2-8khz`	Bilingual with English (code-switching)
Hinglish	`hi-en`	`hi-en-banking-v2-8khz`	Bilingual with English, retained for compatibility
Kannada	`kn`	`kn-banking-v2-8khz`	Bilingual with English (code-switching)
Malayalam	`ml`	`ml-banking-v2-8khz`	Bilingual with English (code-switching)
Marathi	`mr`	`mr-banking-v2-8khz`	Bilingual with English (code-switching)
Odia	`or`	`or-general-v3-8khz`	Monolingual
Tamil	`ta`	`ta-banking-v2-8khz`	Bilingual with English (code-switching)
Telugu	`te`	`te-banking-v2-8khz`	Bilingual with English (code-switching)

Odia has no banking model yet, so or-general-v3-8khz is its only option today.

Reading a model name

hi-banking-v2-8khz
│   │       │  └── training sample rate
│   │       └───── model version
│   └───────────── domain
└───────────────── language

Odia is the one exception to the version number: its model is or-general-v3-8khz, not v2.

Domain. These models are tuned for financial vocabulary: account numbers, balances, transaction terms, the code-switched English that shows up in Indian banking calls.

Sample rate. Every model is trained at 8 kHz, the rate telephony actually delivers. Feeding higher-rate audio doesn’t improve accuracy on a model trained at 8 kHz, and downsampling a 16 kHz recording to 8 kHz before transcribing usually matches the training distribution better.

Choosing a model

Start from the language

There’s no detection and no fallback. A Hindi model on Tamil audio returns a poor transcript rather than an error, so pick the model from your own metadata about the call.

Validate on your own audio

Run a sample set through the model and check per-segment confidence from aux_info. That’s a cheap, objective way to spot-check without labelled ground truth.

Non-streaming transcription

Transcribe a file with one of these models.

Streaming transcription

Declare a model in the config frame.

API reference

Where the model name goes in each request.