Introduction
1.
Installation
2.
Key Features
3.
Product overview
4.
Performance
5.
ML training flow
5.1.
Installation
5.2.
Model YAML configurations
5.3.
Training times
5.4.
Data preparation
5.4.1.
Supported dataset formats
5.4.1.1.
JSONL format
5.4.1.2.
SHAR format
5.4.2.
Log mel feature normalization
5.5.
Training
5.5.1.
Batch size hyperparameters
5.5.2.
Heterogeneous CPUs
5.5.3.
Tensorboard
5.5.4.
Challenging target data
5.5.5.
Resuming and fine-tuning
5.5.6.
Profiling
5.5.7.
Changing the character set
5.5.8.
Language detection
5.5.9.
Code switching
5.5.10.
Delay penalty
5.5.11.
Weakly supervised transducer
5.5.12.
Endpointing
5.5.13.
Conditional training
5.5.14.
End of input (EOI) token
5.5.15.
Checkpoint averaging
5.6.
Validation
5.6.1.
WER calculation
5.6.2.
PER calculation
5.6.3.
State resets
5.6.4.
Beam search decoding
5.6.5.
Blank bias
5.6.6.
N-gram language model
5.6.7.
Automatic batch size reduction
5.6.8.
Saving predictions
5.6.9.
User-perceived latency
5.6.10.
Emission latency
5.6.11.
Timestamps
5.6.12.
Conditional decoding
5.6.13.
Keyword boosting
5.7.
Export inference checkpoint
5.7.1.
Config & checkpoint hierarchy
6.
Inference flow
6.1.
Licensing
6.2.
Programming the FPGA
6.3.
CAIMAN-ASR server
6.3.1.
Websocket API
6.3.2.
Testing inference performance
6.3.3.
Live audio demonstration
6.3.4.
Hardware requirements
6.4.
CAIMAN-ASR demo
6.5.
CAIMAN-ASR benchmark
6.5.1.
Benchmarking custom datasets
7.
Versions
Light (default)
Rust
Coal
Navy
Ayu
CAIMAN-ASR
Versions
Documentation for previous versions:
v1.9.0
v1.10.0
v1.10.1
v1.11.0
v1.12.0
v1.13.0
v1.14.0
v1.16.0
v1.17.0