Language Model Factory

In order to always offer the best speech recognition tool possible, we have developed the Language Model Factory (LM Factory, or LMF). It allows users without technical background to fine tune their language model, and adapt them for their different use cases.

How does it work

It all starts from the use case we want to address, for example company Acme Corp wants the following:

We want customers to track on their own, by phone, their order status.

The customer experience might consist of:

  • first step, a callbot, during which customers give information about their order, such as product name, tracking number...
  • an optional second step with an agent, in case the callbot does not give enough feedback


Through a website, three corpuses will be created. Each one lists language elements specific to the company, which don't exist in the generic Language Model we provide. Products corpus

10t anvil
25t anvil
earthquake pills
water pistol
Rocket powered roller skates
Jet propelled unicycle
Man-sized slingshot
Invisible ink

Competitors corpus

Will E Coyote Limited
Looney Tunes Factories

Acme Corp's trade lingo corpus

Acme Corp
Welcome to Acme Corp
Please don't be anvil !

Language Models (LM)

A first LM is created in order to deal with everything related to Acme Corp.

  • base: generic LM provided by Allo-Media
  • additional corpuses: the ones listed above

A second LM is created in order for the callbot to ask about the product.

  • base: generic LM provided by Allo-Media
  • additional corpus: the one about products

Why a second LM? By focusing only on the products, we increase statistical chances of having a correct transcript on this theme.


A customer call the phone number, and navigate to the tracking order part. The bot asks for the order number, thanks to Allo-Media's own capacities to recognize spelling (thanks to MRCP or WS protocols). The bot also asks some additional question about a product, thanks to the second LM that was created. If necessary, the customer talks to an agent, and the conversation is transcribed and analyzed thanks to the first LM.