NeMo has scripts to convert several common ASR datasets into the format expected by the nemo_asr collection. You can get started with those datasets by following the instructions to run those scripts in the section appropriate to each dataset below.

If the user has their own data and want to preprocess it to use with NeMo ASR models, refer to the Preparing Custom ASR Data section.

If the user already has a dataset that you want to convert to a tarred format, refer to the Tarred Datasets section.


Run the following scripts to download the LibriSpeech data and convert it into the format expected by nemo_asr. At least 250GB free space is required.

# install sox
sudo apt-get install sox
mkdir data
python --data_root=data --data_set=ALL

After this, the data folder should contain wav files and .json manifests for NeMo ASR datalayer.

Each line is a training example. audio_filepath contains the path to the wav file, duration is the duration in seconds, and text is the transcript:

{"audio_filepath": "<absolute_path_to>/1355-39947-0000.wav", "duration": 11.3, "text": "psychotherapy and the community both the physician and the patient find their place in the community the life interests of which are superior to the interests of the individual"}
{"audio_filepath": "<absolute_path_to>/1355-39947-0001.wav", "duration": 15.905, "text": "it is an unavoidable question how far from the higher point of view of the social mind the psychotherapeutic efforts should be encouraged or suppressed are there any conditions which suggest suspicion of or direct opposition to such curative work"}

Fisher English Training Speech

Run these scripts to convert the Fisher English Training Speech data into a format expected by the nemo_asr collection.

In brief, the following scripts convert the .sph files to .wav, slices those files into smaller audio samples, matches the smaller slices with their corresponding transcripts, and splits the resulting audio segments into train, validation, and test sets (with one manifest each).


  • 106 GB of space is required to run the .wav conversion

  • additional 105 GB is required for the slicing and matching

  • sph2pipe is required in order to run the .wav conversion


The following scripts assume that you already have the Fisher dataset from the Linguistic Data Consortium, with a directory structure that looks similar to the following:

├── LDC2004S13-Part1
│   ├── fe_03_p1_transcripts
│   ├── fisher_eng_tr_sp_d1
│   ├── fisher_eng_tr_sp_d2
│   ├── fisher_eng_tr_sp_d3
│   └── ...
└── LDC2005S13-Part2
    ├── fe_03_p2_transcripts
    ├── fe_03_p2_sph1
    ├── fe_03_p2_sph2
    ├── fe_03_p2_sph3
    └── ...

The transcripts that will be used are located in the fe_03_p<1,2>_transcripts/data/trans directory. The audio files (.sph) are located in the remaining directories in an audio subdirectory.

  1. Convert the audio files from .sph to .wav by running:

    cd <nemo_root>/scripts/dataset_processing
    python \
      --data_root=<fisher_root> --dest_root=<conversion_target_dir>

    This will place the unsliced .wav files in <conversion_target_dir>/LDC200[4,5]S13-Part[1,2]/audio-wav/. It will take several minutes to run.

  2. Process the transcripts and slice the audio data.

    python \
      --audio_root=<conversion_target_dir> --transcript_root=<fisher_root> \
      --dest_root=<processing_target_dir> \

    This script splits the full dataset into train, validation, test sets, and places the audio slices in the corresponding folders in the destination directory. One manifest is written out per set, which includes each slice’s transcript, duration, and path.

    This will likely take around 20 minutes to run. Once finished, delete the 10 minute long .wav files.

2000 HUB5 English Evaluation Speech

Run the following script to convert the HUB5 data into a format expected by the nemo_asr collection.

Similarly, to the Fisher dataset processing scripts, this script converts the .sph files to .wav, slices the audio files and transcripts into utterances, and combines them into segments of some minimum length (default is 10 seconds). The resulting segments are all written out to an audio directory and the corresponding transcripts are written to a manifest JSON file.


  • 5 GB of free space is required to run this script

  • sph2pipe is also required to be installed

This script assumes you already have the 2000 HUB5 dataset from the Linguistic Data Consortium.

Run the following command to process the 2000 HUB5 English Evaluation Speech samples:

python \
  --data_root=<path_to_HUB5_data> \

You can optionally include --min_slice_duration=<num_seconds> if you would like to change the minimum audio segment duration.

AN4 Dataset

This is a small dataset recorded and distributed by Carnegie Mellon University. It consists of recordings of people spelling out addresses, names, etc. Information about this dataset can be found on the official CMU site.

  1. Download and extract the dataset (which is labeled “NIST’s Sphere audio (.sph) format (64M)”.

  2. Convert the .sph files to .wav using sox, and build one training and one test manifest.

    python --data_root=<path_to_extracted_data>

After the script finishes, the train_manifest.json and test_manifest.json can be found in the <data_root>/an4/ directory.


To download the Aishell-1 data and convert it into a format expected by nemo_asr, run:

# install sox
sudo apt-get install sox
mkdir data
python --data_root=data

After the script finishes, the data folder should contain a data_aishell folder which contains a wav file, a transcript folder, and related .json and vocab.txt files.


To process the AIShell-2 dataset, in the command below, set the data folder of AIShell-2 using --audio_folder and where to push these files using --dest_folder. In order to generate files in the supported format of nemo_asr, run:

python --audio_folder=<data directory> --dest_folder=<destination directory>

After the script finishes, the train.json, dev.json, test.json, and vocab.txt files can be found in the dest_folder directory.

Preparing Custom ASR Data

The nemo_asr collection expects each dataset to consist of a set of utterances in individual audio files plus a manifest that describes the dataset, with information about one utterance per line (.json). The audio files can be of any format supported by Pydub, though we recommend WAV files as they are the default and have been most thoroughly tested.

There should be one manifest file per dataset that will be passed in, therefore, if the user wants separate training and validation datasets, they should also have separate manifests. Otherwise, they will be loading validation data with their training data and vice versa.

Each line of the manifest should be in the following format:

{"audio_filepath": "/path/to/audio.wav", "text": "the transcription of the utterance", "duration": 23.147}

The audio_filepath field should provide an absolute path to the .wav file corresponding to the utterance. The text field should contain the full transcript for the utterance, and the duration field should reflect the duration of the utterance in seconds.

Each entry in the manifest (describing one audio file) should be bordered by ‘{’ and ‘}’ and must be contained on one line. The fields that describe the file should be separated by commas, and have the form "field_name": value, as shown above. There should be no extra lines in the manifest, i.e. there should be exactly as many lines in the manifest as there are audio files in the dataset.

Since the manifest specifies the path for each utterance, the audio files do not have to be located in the same directory as the manifest, or even in any specific directory structure.

Once there is a manifest that describes each audio file in the dataset, use the dataset by passing in the manifest file path in the experiment config file, e.g. as training_ds.manifest_filepath=<path/to/manifest.json>.

Tarred Datasets

If experiments are run on a cluster with datasets stored on a distributed file system, the user will likely want to avoid constantly reading multiple small files and would prefer tarring their audio files. There are tarred versions of some NeMo ASR dataset classes for this case, such as the TarredAudioToCharDataset (corresponding to the AudioToCharDataset) and the TarredAudioToBPEDataset (corresponding to the AudioToBPEDataset). The tarred audio dataset classes in NeMo use WebDataset.

To use an existing tarred dataset instead of a non-tarred dataset, set is_tarred: true in the experiment config file. Then, pass in the paths to all of the audio tarballs in tarred_audio_filepaths, either as a list of filepaths, e.g. ['/data/shard1.tar', '/data/shard2.tar'], or in a single brace-expandable string, e.g. '/data/shard_{1..64}.tar' or '/data/shard__OP_1..64_CL_' (recommended, see note below).


For brace expansion, there may be cases where {x..y} syntax cannot be used due to shell interference. This occurs most commonly inside SLURM scripts. Therefore, we provide a few equivalent replacements. Supported opening braces (equivalent to {) are (, [, < and the special tag _OP_. Supported closing braces (equivalent to }) are ), ], > and the special tag _CL_. For SLURM based tasks, we suggest the use of the special tags for ease of use.

As with non-tarred datasets, the manifest file should be passed in manifest_filepath. The dataloader assumes that the length of the manifest after filtering is the correct size of the dataset for reporting training progress.

The tarred_shard_strategy field of the config file can be set if you have multiple shards and are running an experiment with multiple workers. It defaults to scatter, which preallocates a set of shards per worker which do not change during runtime. Note that this strategy, on specific occasions (when the number of shards is not divisible with world_size), will not sample the entire dataset. As an alternative the replicate strategy, will preallocate the entire set of shards to every worker and not change it during runtime. The benefit of this strategy is that it allows each worker to sample data points from the entire dataset independently of others. Note, though, that more than one worker may sample the same shard, and even sample the same data points! As such, there is no assured guarantee that all samples in the dataset will be sampled at least once during 1 epoch. Note that for these reasons it is not advisable to use tarred datasets as validation and test datasets.

For more information about the individual tarred datasets and the parameters available, including shuffling options, see the corresponding class APIs in the Datasets section.


If using multiple workers, the number of shards should be divisible by the world size to ensure an even split among workers. If it is not divisible, logging will give a warning but training will proceed, but likely hang at the last epoch. In addition, if using distributed processing, each shard must have the same number of entries after filtering is applied such that each worker ends up with the same number of files. We currently do not check for this in any dataloader, but the user’s program may hang if the shards are uneven.

Sharded Manifests

If your dataset / manifest is large, you may wish to use sharded manifest files instead of a single manifest file. The naming convention is identical to the audio tarballs and there should be a 1:1 relationship between a sharded audio tarfile and its manifest shard; e.g. '/data/sharded_manifests/manifest__OP_1..64_CL_' in the above example. Using sharded manifests improves job startup times and decreases memory usage, as each worker only loads manifest shards for the corresponding audio shards instead of the entire manifest.

To enable sharded manifest filename expansion, set the shard_manifests field of the config file to true. In addition, the defer_setup flag needs to be true as well, so that the dataloader will be initialized after the DDP and its length can be collected from the distributed workers.

Batching strategies

For training ASR models, audios with different lengths may be grouped into a batch. It would make it necessary to use paddings to make all the same length. These extra paddings is a significant source of computation waste.

Semi Sorted Batching

Sorting samples by duration and spliting them into batches speeds up training, but can degrade the quality of the model. To avoid quality degradation and maintain some randomness in the partitioning process, we add pseudo noise to the sample length when sorting.

It may result into training speeedup of more than 40 percent with the same quality. To enable and use semi sorted batching add some lines in config.


Semi sorted batching is supported by the following models:


For more details about this algorithm, see the paper .

Bucketing Datasets

Splitting the training samples into buckets with different lengths and sampling from the same bucket for each batch would increase the computation efficicncy. It may result into training speeedup of more than 2X. To enable and use the bucketing feature, you need to create the bucketing version of the dataset by using conversion script here. You may use –buckets_num to specify the number of buckets (Recommend to use 4 to 8 buckets). It creates multiple tarred datasets, one per bucket, based on the audio durations. The range of [min_duration, max_duration) is split into equal sized buckets.

To enable the bucketing feature in the dataset section of the config files, you need to pass the multiple tarred datasets as a list of lists. If user passes just a list of strings, then the datasets would simply get concatenated which would be different from bucketing. Here is an example for 4 buckets and 512 shards:


When bucketing is enabled, in each epoch, first all GPUs would use the first bucket, then go to the second bucket, and so on. It guarantees that all GPUs are using the same bucket at the same time. It reduces the number of paddings in each batch and speedup the training significantly without hurting the accuracy significantly.

There are two types of batching:

  • Fixed-size bucketing: all batches would have the same number of samples specified by train_ds.batch_size

  • Adaptive-size bucketing: uses different batch sizes for each bucket.

Adaptive-size bucketing helps to increase the GPU utilization and speedup the training. Batches sampled from buckets with smaller audio lengths can be larger which would increase the GPU utilization and speedup the training. You may use train_ds.bucketing_batch_size to enable the adaptive batching and specify the batch sizes for the buckets. When bucketing_batch_size is not set, train_ds.batch_size is going to be used for all buckets (fixed-size bucketing).

bucketing_batch_size can be set as an integer or a list of integers to explicitly specify the batch size for each bucket. if bucketing_batch_size is set to be an integer, then linear scaling is being used to scale-up the batch sizes for batches with shorted audio size. For example, setting train_ds.bucketing_batch_size=8 for 4 buckets would use these sizes [32,24,16,8] for different buckets. When bucketing_batch_size is set, traind_ds.batch_size need to be set to 1.

Training an ASR model on audios sorted based on length may affect the accuracy of the model. We introduced some strategies to mitigate it. We support three types of bucketing strategies:

  • fixed_order: the same order of buckets are used for all epochs

  • synced_randomized (default): each epoch would have a different order of buckets. Order of the buckets is shuffled every epoch.

  • fully_randomized: similar to synced_randomized but each GPU has its own random order. So GPUs would not be synced.

Tha parameter train_ds.bucketing_strategy can be set to specify one of these strategies. The recommended strategy is synced_randomized which gives the highest training speedup. The fully_randomized strategy would have lower speedup than synced_randomized but may give better accuracy.

Bucketing may improve the training speed more than 2x but may affect the final accuracy of the model slightly. Training for more epochs and using ‘synced_randomized’ strategy help to fill this gap. Currently bucketing feature is just supported for tarred datasets.

Conversion to Tarred Datasets

You can easily convert your existing NeMo-compatible ASR datasets using the conversion script here.

python \
  --manifest_path=<path to the manifest file> \
  --target_dir=<path to output directory> \
  --num_shards=<number of tarfiles that will contain the audio>
  --max_duration=<float representing maximum duration of audio samples> \
  --min_duration=<float representing minimum duration of audio samples> \
  --shuffle --shuffle_seed=0

This script shuffles the entries in the given manifest (if --shuffle is set, which we recommend), filter audio files according to min_duration and max_duration, and tar the remaining audio files to the directory --target_dir in n shards, along with separate manifest and metadata files.

The files in the target directory should look similar to the following:

├── audio_1.tar
├── audio_2.tar
├── ...
├── metadata.yaml
├── tarred_audio_manifest.json
├── sharded_manifests/
    ├── manifest_1.json
    ├── ...
    └── manifest_N.json

Note that file structures are flattened such that all audio files are at the top level in each tarball. This ensures that filenames are unique in the tarred dataset and the filepaths do not contain “-sub” and forward slashes in each audio_filepath are simply converted to underscores. For example, a manifest entry for /data/directory1/file.wav would be _data_directory1_file.wav in the tarred dataset manifest, and /data/directory2/file.wav would be converted to _data_directory2_file.wav.

Sharded manifests are generated by default; this behavior can be toggled via the no_shard_manifests flag.

Upsampling Datasets

Buckets may also be ‘weighted’ to allow multiple runs through a target dataset during each training epoch. This can be beneficial in cases when a dataset is composed of several component sets of unequal sizes and one desires to mitigate bias towards the larger sets through oversampling.

Weighting is managed with the bucketing_weights parameter. After passing your composite tarred datasets in the format described above for bucketing, pass a list of integers (one per bucket) to indicate how many times a manifest should be read during training.

For example, by passing [2,1,1,3] to the code below:


NeMo will configure training so that all data in bucket1 will be present twice in a training epoch, bucket4 will be present three times, and that of bucket2 and bucket3 will occur only once each. Note that this will increase the effective amount of data present during training and thus affect training time per epoch.

If using adaptive bucketing, note that the same batch size will be assigned to each instance of the upsampled data. That is, given the following:


All instances of data from bucket4 will still be trained with a batch size of 2 while all others would have a batch size of 4. As with standard bucketing, this requires batch_size` to be set to 1. If bucketing_batch_size is not specified, all datasets will be passed with the same fixed batch size as specified by the batch_size parameter.

It is recommended to set bucketing strategies to fully_randomized during multi-GPU training to prevent possible dataset bias during training.

Datasets on AIStore

AIStore is an open-source lightweight object storage system focused on large-scale deep learning. AIStore is aimed to scale linearly with each added storage node, can be deployed on any Linux machine and can provide a unified namespace across multiple remote backends, such as Amazon S3, Google Cloud, and Microsoft Azure. More details are provided in the documentation and the repository of the AIStore project.

NeMo currently supports datasets from an AIStore bucket provider under ais:// namespace.

AIStore Setup

NeMo is currently relying on the AIStore (AIS) command-line interface (CLI) to handle the supported datasets. The CLI is available in current NeMo Docker containers. If necessary, the CLI can be configured using the instructions provided in AIStore CLI documentation.

To start using the AIS CLI to access data on an AIS cluster, an endpoint needs to be configured. The endpoint is configured by setting AIS_ENDPOINT environment variable before using the CLI

export AIS_ENDPOINT=http://hostname:port
ais --help

In the above, hostname:port denotes the address of an AIS gateway. For example, the address could be localhost:51080 if testing using a local minimal production-ready standalone Docker container.

Dataset Setup

Currently, both tarred and non-tarred datasets are supported. For any dataset, the corresponding manifest file is cached locally and processed as a regular manifest file. For non-tarred datasets, the audio data is also cached locally. For tarred datasets, shards from the AIS cluster are used by piping ais get to WebDataset.

Tarred Dataset from AIS

A tarred dataset can be easily used as described in the Tarred Datasets section by providing paths to manifests on an AIS cluster. For example, a tarred dataset from an AIS cluster can be configured as


Bucketing Datasets are configured in a similar way by providing paths on an AIS cluster.

Non-tarred Dataset from AIS

A non-tarred dataset can be easly used by providing a manifest file path on an AIS cluster


Note that it is assumed that the manifest file path contains audio file paths relative to the manifest locations. For example the manifest file may have lines in the following format

{"audio_filepath": "path/to/audio.wav", "text": "transcription of the uterance", "duration": 23.147}

The corresponding audio file would be downloaded from ais://bucket/path/to/audio.wav.

Cache configuration

Manifests and audio files from non-tarred datasets will be cached locally. Location of the cache can be configured by setting two environment variables

  • NEMO_DATA_STORE_CACHE_DIR: path to a location which can be used to cache the data

  • NEMO_DATA_STORE_CACHE_SHARED: flag to denote whether the cache location is shared between the compute nodes

In a multi-node environment, the cache location may or may be not shared between the nodes. This can be configured by setting NEMO_DATA_STORE_CACHE_SHARED to 1 when the location is shared between the nodes or to 0 when each node has a separate cache.

When a globally shared cache is available, the data should be cached only once from the global rank zero node. When a node-specific cache is used, the data should be cached only once by each local rank zero node. To control this behavior using torch.distributed.barrier, instantiation of the corresponding dataloader needs to be deferred ModelPT::setup, to ensure a distributed environment has been initialized. This can be achieved by setting defer_setup as


Complete Example

An example using an AIS cluster at hostname:port with a tarred dataset for training, a non-tarred dataset for validation and node-specific caching is given below

export AIS_ENDPOINT=http://hostname:port \
&& export NEMO_DATA_STORE_CACHE_DIR=/tmp \
python \
model.train_ds.manifest_filepath=ais://train_bucket/tarred_audio_manifest.json \
model.train_ds.tarred_audio_filepaths=ais://train_bucket/audio__OP_0..511_CL_.tar \
++model.train_ds.defer_setup=true \
mode.validation_ds.manifest_filepath=ais://validation_bucket/validation_manifest.json \

Lhotse Dataloading

NeMo supports using Lhotse, a speech data handling library, as a dataloading option. The key features of Lhotse used in NeMo are:

  • Dynamic batch sizes

    Lhotse samples mini-batches to satisfy the constraint of total speech duration in a mini-batch (batch_duration), rather than a specific number of examples (i.e., batch size).

  • Dynamic bucketing

    Instead of statically pre-bucketing the data, Lhotse allocates training examples to buckets dynamically. This allows more rapid experimentation with bucketing settings (number of buckets, specific placement of bucket duration bins) to minimize the amount of padding and accelerate training.

  • Quadratic duration penalty

    Adding a quadratic penalty to an utterance’s duration allows to sample mini-batches so that the GPU utilization is more consistent across big batches of short utterances and small batches of long utterances when using models with quadratic time/memory complexity (such as transformer).

  • Dynamic weighted data source multiplexing

    An approach to combining diverse data sources (e.g. multiple domains, languages, tasks) where each data source is treated as a separate stream with its own sampling probability. The resulting data stream is a multiplexer that samples from each sub-stream. This approach ensures that the distribution of different sources is approximately constant in time (i.e., stationary); in fact, each mini-batch will have roughly the same ratio of data coming from each source. Since the multiplexing is done dynamically, it is very easy to tune the sampling weights.

Lhotse dataloading supports the following types of inputs:

  • NeMo manifests

    Regular NeMo JSON manifests.

  • NeMo tarred data

    Tarred NeMo JSON manifests + audio tar files; we also support combination of multiple NeMo tarred data sources (e.g., multiple buckets of NeMo data or multiple datasets) via dynamic multiplexing.

  • Lhotse CutSet manifests

    Regular Lhotse CutSet manifests (typically gzipped JSONL). See Lhotse Cuts documentation to learn more about Lhotse data formats.

  • Lhotse Shar data

    Lhotse Shar is a data format that also uses tar files for sequential data loading, but is designed to be modular (i.e., easily extensible with new data sources and with new feature fields). More details can be found here: tutorial_shar


As of now, Lhotse is mainly supported in most ASR model configurations. We aim to gradually extend this support to other speech tasks.

Enabling Lhotse via configuration


Using Lhotse with tarred datasets will make the dataloader infinite, ditching the notion of an “epoch”. “Epoch” may still be logged in W&B/TensorBoard, but it will correspond to the number of executed training loops between validation loops.

Start with an existing NeMo experiment YAML configuration. Typically, you’ll only need to add a few options to enable Lhotse. These options are:

# NeMo generic dataloading arguments
model.train_ds.tarred_audio_filepaths=...   # for tarred datasets only
model.train_ds.min_duration=0.3             # optional
model.train_ds.max_duration=30.0            # optional
model.train_ds.shuffle=true                 # optional

# Lhotse dataloading related arguments

# PyTorch Lightning related arguments


The default values above are a reasonable starting point for a hybrid RNN-T + CTC ASR model on a 32GB GPU with a data distribution dominated by 15s long utterances.

Let’s briefly go over each of the Lhotse dataloading arguments:

  • use_lhotse enables Lhotse dataloading

  • batch_duration is the total max duration of utterances in a mini-batch and controls the batch size; the more shorter utterances, the bigger the batch size, and vice versa.

  • quadratic_duration adds a quadratically growing penalty for long utterances; useful in bucketing and transformer type of models. The value set here means utterances this long will count as if with a doubled duration.

  • num_buckets is the number of buckets in the bucketing sampler. Bigger value means less padding but also less randomization.

  • num_cuts_for_bins_estimate is the number of utterance we will sample before the start of the training to estimate the duration bins for buckets. Larger number results in a more accurate estimatation but also a bigger lag before starting the training.

  • bucket_buffer_size is the number of utterances (data and metadata) we will hold in memory to be distributed between buckets. With bigger batch_duration, this number may need to be increased for dynamic bucketing sampler to work properly (typically it will emit a warning if this is too low).

  • shuffle_buffer_size is an extra number of utterances we will hold in memory to perform approximate shuffling (via reservoir-like sampling). Bigger number means more memory usage but also better randomness.

The PyTorch Lightning trainer related arguments:

  • use_distributed_sampler=false is required because Lhotse has its own handling of distributed sampling.

  • val_check_interval/limit_train_batches

    These are required for dataloaders with tarred/Shar datasets because Lhotse makes the dataloader infinite, so we’d never go past epoch 0. This approach guarantees we will never hang the training because the dataloader in some node has less mini-batches than the others in some epochs. The value provided here will be the effective length of each “pseudo-epoch” after which we’ll trigger the validation loop.

  • max_steps is the total number of steps we expect to be training for. It is required for the same reason as limit_train_batches; since we’d never go past epoch 0, the training would have never finished.

Some other Lhotse related arguments we support:

  • cuts_path can be provided to read data from a Lhotse CutSet manifest instead of a NeMo manifest.

    Specifying this option will result in manifest_filepaths and tarred_audio_filepaths being ignored.

  • shar_path

    Can be provided to read data from a Lhotse Shar manifest instead of a NeMo manifest. This argument can be a string (single Shar directory), a list of strings (Shar directories), or a list of 2-item lists, where the first item is a Shar directory path, and the other is a sampling weight. Specifying this option will result in manifest_filepaths and tarred_audio_filepaths being ignored.

  • bucket_duration_bins

    Duration bins are a list of float values (seconds) that when provided, will skip the initial bucket bin estimation and save some time. It has to have a length of num_buckets - 1. An optimal value can be obtained by running CLI: lhotse cut estimate-bucket-bins -b $num_buckets my-cuts.jsonl.gz

  • use_bucketing is a boolean which indicates if we want to enable/disable dynamic bucketing. By defalt it’s enabled.

  • text_field is the name of the key in the JSON (NeMo) manifest from which we should be reading text (default=”text”).

  • lang_field is the name of the key in the JSON (NeMo) manifest from which we should be reading language tag (default=”lang”). This is useful when working e.g. with AggregateTokenizer.

  • batch_size

    Limits the number of examples in a mini-batch to this number, when combined with batch_duration. When batch_duration is not set, it acts as a static batch size.

  • seed sets a random seed for the shuffle buffer.

The full and always up-to-date list of supported options can be found in LhotseDataLoadingConfig class.

Extended multi-dataset configuration format

Combining a large number of datasets and defining weights for them can be tricky. We offer an extended configuration format that allows you to explicitly define datasets, dataset groups, and their weights either inline in the experiment configuration, or as a path to a separate YAML file.

In addition to the features above, this format introduces a special tags dict-like field. The keys and values in tags are automatically attached to every sampled example, which is very useful when combining multiple datasets with different properties. The dataset class which converts these examples to tensors can partition the mini-batch and apply different processing to each group. For example, you may want to construct different prompts for the model using metadata in tags.


When fine-tuning a model that was trained with input_cfg option, typically you’d only need to override the following options: input_cfg=null and manifest_filepath=path/to/manifest.json.

Example 1. Combine two datasets with equal weights and attach custom metadata in tags to each cut:

  - type: nemo_tarred
    manifest_filepath: /path/to/manifest__OP_0..512_CL_.json
    tarred_audio_filepath: /path/to/tarred_audio/audio__OP_0..512_CL_.tar
    weight: 0.4
      lang: en
      pnc: no
  - type: nemo_tarred
    manifest_filepath: /path/to/other/manifest__OP_0..512_CL_.json
    tarred_audio_filepath: /path/to/other/tarred_audio/audio__OP_0..512_CL_.tar
    weight: 0.6
      lang: pl
      pnc: yes

Example 2. Combine multiple (4) datasets, corresponding to different tasks (ASR, AST). Each task gets its own group and its own weight. Then within each task, each dataset get its own within-group weight as well. The final weight is the product of outer and inner weight:

  - type: group
    weight: 0.7
      task: asr
      - type: nemo_tarred
        manifest_filepath: /path/to/asr1/manifest__OP_0..512_CL_.json
        tarred_audio_filepath: /path/to/tarred_audio/asr1/audio__OP_0..512_CL_.tar
        weight: 0.6
          source_lang: en
          target_lang: en
      - type: nemo_tarred
        manifest_filepath: /path/to/asr2/manifest__OP_0..512_CL_.json
        tarred_audio_filepath: /path/to/asr2/tarred_audio/audio__OP_0..512_CL_.tar
        weight: 0.4
          source_lang: pl
          target_lang: pl
  - type: group
    weight: 0.3
      task: ast
      - type: nemo_tarred
        manifest_filepath: /path/to/ast1/manifest__OP_0..512_CL_.json
        tarred_audio_filepath: /path/to/ast1/tarred_audio/audio__OP_0..512_CL_.tar
        weight: 0.2
          source_lang: en
          target_lang: pl
      - type: nemo_tarred
        manifest_filepath: /path/to/ast2/manifest__OP_0..512_CL_.json
        tarred_audio_filepath: /path/to/ast2/tarred_audio/audio__OP_0..512_CL_.tar
        weight: 0.8
          source_lang: pl
          target_lang: en

Configuring multi-modal dataloading

Our configuration format supports specifying data sources from other modalities than just audio. At this time, this support is extended to text-only data. We provide the following parser types:

  • txt for raw text files, sharded or unsharded. This can represent, for example, language modeling data.

  • txt_pair for pairs of raw text files, sharded or unsharded. This can represent, for example, machine translation data.

The key strength of this approach is that we can easily combine audio datasets and text datasets, and benefit from every other technique we described above such as dynamic data mixing, data weighting, dynamic bucketing, and so on. To enable multimodal dataloading, we provide several configuration options:

  • use_multimodal_sampling when set to True, we’ll discard the settings of batch_duration and quadratic_duration and consider the settings below instead.

  • batch_tokens is the maximum number of tokens we want to find inside a mini-batch. Similarly to batch_duration, this number does consider padding tokens too, therefore enabling bucketing is recommended to maximize the ratio of real vs padding tokens.

  • token_equivalent_duration is used to be able to measure audio examples in the number of “tokens”. For example, if we’re using fbank with 0.01s frame shift and an acoustic model that has a subsampling factor of 0.08, then a reasonable setting for this could be 0.08 (which means every subsampled frame counts as one token). Calibrate this value to fit your needs. Note that this value acts as a “balancer” between how much audio data vs text data gets sampled into a mini-batch.

  • quadratic_factor works the same way as quadratic_duration, but is defined in the number of tokens.

Example 3. Combine an ASR (audio-text) dataset with an MT (text-only) dataset so that mini-batches have some examples from both datasets. Provide a custom prompt field for both datasets (to be leveraged by a relevant dataset class):

use_multimodal_sampling: true
batch_tokens: 1024
token_equivalent_duration: 0.08  # 0.01 frame shift * 8 subsampling factor
quadratic_factor: 50
num_buckets: 30
use_bucketing: true
  - type: nemo_tarred
    manifest_filepath: /path/to/manifest__OP_0..512_CL_.json
    tarred_audio_filepath: /path/to/tarred_audio/audio__OP_0..512_CL_.tar
    weight: 0.5
      lang: en
      prompt: "Given the following recording, transcribe what the person is saying:"
  - type: txt_pair
    source_path: /path/to/en__OP_0..512_CL_.txt
    target_path: /path/to/pl__OP_0..512_CL_.txt
    source_language: en
    target_language: pl
    weight: 0.5
      prompt: "Translate the following text to Polish:"


We strongly recommend to use multiple shards for text files as well so that different nodes and dataloading workers are able to randomize the order of text iteration. Otherwise, multi-GPU training has a high risk of duplication of text examples.

Pre-computing bucket duration bins

We recommend to pre-compute the bucket duration bins in order to accelerate the start of the training – otherwise, the dynamic bucketing sampler will have to spend some time estimating them before the training starts. The following script may be used:

$ python scripts/speech_recognition/ -b 30 manifest.json
Use the following options in your config:
<other diagnostic information about the dataset>

For multi-dataset setups, one may provide multiple manifests and even their weights:

$ python scripts/speech_recognition/ -b 30 [[manifest.json,0.7],[other.json,0.3]]
Use the following options in your config:
<other diagnostic information about the dataset>

Seeds and randomness

In Lhotse dataloading configuration we have two parameters controlling randomness: seed and shard_seed. Both of them can be either set to a fixed number, or one of two string options "randomized" and "trng". Their roles are:

  • seed is the base random seed, and is one of several factors used to initialize various RNGs participating in dataloading.

  • shard_seed controls the shard randomization strategy in distributed data parallel setups when using sharded tarred datasets.

Below are the typical examples of configuration with an explanation of the expected outcome.

Case 1 (default): seed=<int> and shard_seed="trng":

  • The trng setting discards seed and causes the actual random seed to be drawn using OS’s true RNG. Each node/GPU/dataloading worker draws its own unique random seed when it first needs it.

  • Each node/GPU/dataloading worker yields data in a different order (no mini-batch duplication).

  • On each training script run, the order of dataloader examples are different.

  • Since the random seed is unpredictable, the exact dataloading order is not replicable.

Case 2: seed=<int> and shard_seed="randomized":

  • The randomized setting uses seed along with DDP rank and dataloading worker_id to set a unique but deterministic random seed in each dataloading process across all GPUs.

  • Each node/GPU/dataloading worker yields data in a different order (no mini-batch duplication).

  • On each training script run, the order of dataloader examples are identical as long as seed is the same.

  • This setup guarantees 100% dataloading reproducibility.

  • Resuming training without changing of the seed value will cause the model to train on data it has already seen. For large data setups, not managing the seed may cause the model to never be trained on a majority of data. This is why this mode is not the default.

  • If you’re combining DDP with model parallelism techniques (Tensor Parallel, Pipeline Parallel, etc.) you need to use shard_seed="randomized". Using "trng" will cause different model parallel ranks to desynchronize and cause a deadlock.

  • Generally the seed can be managed by the user by providing a different value each time the training script is launched. For example, for most models the option to override would be model.train_ds.seed=<value>. If you’re launching multiple tasks queued one after another on a grid system, you can generate a different random seed for each task, e.g. on most Unix systems RSEED=$(od -An -N4 -tu4 < /dev/urandom | tr -d ' ') would generate a random uint32 number that can be provided as the seed.

Other, more exotic configurations:

  • With shard_seed=<int>, all dataloading workers will yield the same results. This is only useful for unit testing and maybe debugging.

  • With seed="trng", the base random seed itself will be drawn using a TRNG. It will be different on each GPU training process. This setting is not recommended.

  • With seed="randomized", the base random seed is set to Python’s global RNG seed. It might be different on each GPU training process. This setting is not recommended.

Preparing Text-Only Data for Hybrid ASR-TTS Models

Hybrid ASR-TTS models require a text-only dataset for training the ASR model. Each record in the dataset (in .json file) should contain the following fields:

  • text: text to use as a target for the ASR model

  • tts_text or/and tts_text_normalized: text to use as a source for TTS model. tts_text_normalized should contain normalized text for TTS model. If there is no such field, tts_text will be used after normalization using the normalizer from the TTS model. It is highly recommended to normalize the text and create tts_text_normalized field manually, since current normalizers are unsuitable for processing a large amount of text on the fly.

Example record:

{"text": "target for one hundred billion parameters asr model",
 "tts_text": "Target for 100B parameters ASR model.",
 "tts_text_normalized": "Target for one hundred billion parameters ASR model."}