Speech Recognition: English N-Gram Language Models

Model Overview

When deployed, the ASR engine can optionally condition the transcript output on n-gram language models.

Model Architecture

These models are simple 3- and 4-gram language models trained with Kneser-Ney smoothing using KenLM.

Intended Use

Primary use case intended for these models is automatic speech recognition.

Input

Sequence of zero or more words.

Output

Likelihood of word sequence.

How to Use this Model

There are a variety of formats contained within this model archive:

ARPA-formatted Language Models: -en-US_default_6.0.arpa

KenLM-formatted Binary Language Models

• en-US_default_6.0.bin

Rescoring Language Models

• G.mixed_lm.3-gram.pruned.3e-7.carpa
• G.mixed_lm.carpa

FST-formatted Language Models

• TLG.mixed_lm.3-gram.pruned.3e-7.fst

Vocabulary Files

• words.mixed_lm.3-gram.pruned.3e-7.txt

Flashlight Decoder Vocabulary Files

• en-US_default_6.0_dict_vocab.txt

ARPA and KenLM binary formatted files can be used directly by the CTC CPU Decoder. The GPU decoder uses a FST-formatted language model (derived from the pruned n-gram model) and then optionally uses the carpa-formatted LMs for rescoring.

Training

The mixed language model provided here is English-only and is trained on a mix of transcriptions from LibriSpeech, Switchboard, and Fisher datasets.

Limitations

Currently, TLT cannot train LMs for ASR inference. To train custom LMs for ASR inference, use KenLM and consult the Riva Documentation.

License

By downloading and using the models and resources packaged with Riva Conversational AI, you accept the terms of the Riva license.

Ethical AI

NVIDIA’s platforms and application frameworks enable developers to build a wide array of AI applications. Consider potential algorithmic bias when choosing or creating the models being deployed. Work with the model’s developer to ensure that it meets the requirements for the relevant industry and use case; that the necessary instruction and documentation are provided to understand error rates, confidence intervals, and results; and that the model is being used under the conditions and in the manner intended.