Why building a new scorer outputs an empty string for deepspeech 0.9.3

Question

I am trying to create a limited corpus and train a language model to use for a deepspeech scorer.

I have followed the information provided in the docs here

I read a helpful guide posted for an older version of deepspeech for generating a language model here

And I have read the playbook here,

It seems that this has been encountered before, but no answer was given there

I have set up the docker environment for training and followed the docs to the letter.

I can train a model, and then convert this to a .scorer file, so the whole process is working.

The steps I take from inside the docker container are:

• create a vocab.txt file with my input sentences and store it in the deepspeech-data-input folder.
• run this script to build the model in the outputpython3 generate_lm.py --input_txt ../../deepspeech-data/input/vocab.txt --output_dir ../../deepspeech-data/output --top_k 100 --kenlm_bins /DeepSpeech/native_client/kenlm/build/bin/ --arpa_order 5 --max_arpa_memory "85%" --arpa_prune "0|0|0|0" --binary_a_bits 255 --binary_q_bits 8 --binary_type trie --discount_fallback
• run this script to generate the scorer: ./generate_scorer_package --alphabet ../../deepspeech-data/input/alphabet.txt --lm ../../deepspeech-data/output/lm.binary --vocab ../../deepspeech-data/output/vocab-100.txt --package ../../deepspeech-data/output/deepspeech-0.9.3-models.scorer --default_alpha 0.9 --default_beta 0.9 --force_bytes_output_mode 1
• replace the default scorer with my one.
• Run deepspeech

Everything seems to work as it should, no errors or anything, but when doing this deepspeech just detects an empty string. If I use the default scorer I have it working fine, but I need to restrict the vocabulary so that I can just detect a few commands.

I have tried adjusting some of the flags, but I always get the same result.

I am using the --discount_fallback flag as suggested as it is a small corpus

So my question is this. Why would a deepspeech language model/scrorer output an empty string and how can I fix it?

I am running this inside the NodeJS example on github but testing against any of them would work to reproduce. github examples