Home

TensorFlowASR ⚡

Almost State-of-the-art Automatic Speech Recognition in Tensorflow 2

TensorFlowASR implements some automatic speech recognition architectures such as DeepSpeech2, Jasper, RNN Transducer, ContextNet, Conformer, etc. These models can be converted to TFLite to reduce memory and computation for deployment 😄

What's New?

Table of Contents

• What's New?
• Table of Contents
• 😋 Supported Models
  • Baselines
  • Publications
• Installation
• Training & Testing Tutorial
• Features Extraction
• Decoders
• Inference
• Augmentations
• TFLite Convertion
• Pretrained Models
• Corpus Sources
  • English
  • Vietnamese
• How to contribute
• References & Credits
• Contact

😋 Supported Models

Baselines

• Transducer Models (End2end models using RNNT Loss for training, currently supported Conformer, ContextNet, Streaming Transducer)
• CTCModel (End2end models using CTC Loss for training, currently supported DeepSpeech2, Jasper)

Publications

• Conformer Transducer (Reference: https://arxiv.org/abs/2005.08100) See examples/models/transducer/conformer
• Streaming Conformer (Reference: http://arxiv.org/abs/2010.11395) See examples/models/transducer/conformer
• ContextNet (Reference: http://arxiv.org/abs/2005.03191) See examples/models/transducer/contextnet
• RNN Transducer (Reference: https://arxiv.org/abs/1811.06621) See examples/models/transducer/rnnt
• Deep Speech 2 (Reference: https://arxiv.org/abs/1512.02595) See examples/models/ctc/deepspeech2
• Jasper (Reference: https://arxiv.org/abs/1904.03288) See examples/models/ctc/jasper

Installation

For training and testing, you should use git clone for installing necessary packages from other authors (ctc_decoders, rnnt_loss, etc.)

TensorFlowASR uses uv as its package manager and requires python 3.12 or 3.13.

A plain uv sync installs tensorflow + tensorflow-text, which is all that CPU and Apple Silicon need. Accelerators are opt-in extras:

git clone https://github.com/TensorSpeech/TensorFlowASR.git
cd TensorFlowASR
uv sync                 # CPU / Apple Silicon
uv sync --extra cuda    # NVIDIA GPU
uv sync --extra dev     # add the development tooling

Extras are declared in pyproject.toml:

Extra	Contents
(none)	tensorflow + tensorflow-text, works on CPU and Apple Silicon
cuda	tensorflow[and-cuda] for NVIDIA GPUs
dev	pytest, ruff, pre-commit, plotting and export tooling

Run commands inside the environment with uv run, e.g. uv run pytest or uv run tensorflow_asr --help.

Cloud TPU is not an extra:

uv sync && ./scripts/install_tpu.sh

tensorflow-tpu ships its own tensorflow distribution, and tensorflow is a base dependency, so an extra adding it would install both and leave whichever landed last in place. The script uninstalls the stock build first, then installs the TPU one. Re-run it after any later uv sync, which puts the stock tensorflow back.

TPU note: tensorflow-tpu ships its own tensorflow distribution and overwrites the base one. This matches the previous setup.sh tpu behaviour, which uninstalled tensorflow and force-installed tensorflow-tpu over it.

Running in a container

docker-compose up -d

Training & Testing Tutorial

• For training, please read tutorial_training
• For testing, please read tutorial_testing

FYI: Keras builtin training uses infinite dataset, which avoids the potential last partial batch.

See examples for some predefined ASR models and results

Features Extraction

See features_extraction

Decoders

Greedy and beam search decoding for CTC and Transducer models, including the ALSD++ transducer beam search

See decoders

Inference

ASRInference transcribes one stream with either a checkpoint or an exported tflite model, in one pass or streaming chunk by chunk. All sessions on one model share one ASREngine, which loads the model once and decodes many sessions per call, for example one session per websocket in a server

See inferences, and examples/inferences for runnable scripts including a live microphone

Augmentations

See augmentations

TFLite Convertion

After converting to tflite, the tflite model is like a function that transforms directly from an audio signal to text and tokens

See tflite_convertion

Pretrained Models

See the results on each example folder, e.g. ./examples/models//transducer/conformer/results/sentencepiece/README.md

Corpus Sources

English

Name	Source	Hours
LibriSpeech	LibriSpeech	970h
Common Voice	https://commonvoice.mozilla.org	1932h

Vietnamese

Name	Source	Hours
Vivos	https://ailab.hcmus.edu.vn/vivos	15h
InfoRe Technology 1	InfoRe1 (passwd: BroughtToYouByInfoRe)	25h
InfoRe Technology 2 (used in VLSP2019)	InfoRe2 (passwd: BroughtToYouByInfoRe)	415h
VietBud500	https://huggingface.co/datasets/linhtran92/viet_bud500	500h

How to contribute

1. Fork the project
2. Install for development
3. Create a branch
4. Make a pull request to this repo

References & Credits

1. NVIDIA OpenSeq2Seq Toolkit
2. https://github.com/noahchalifour/warp-transducer
3. Sequence Transduction with Recurrent Neural Network
4. End-to-End Speech Processing Toolkit in PyTorch
5. https://github.com/iankur/ContextNet

Contact

Huy Le Nguyen

Email: nlhuy.cs.16@gmail.com