TMS Express 🗣

TMS Express generates bitstreams for the TMS5220 Voice Synthesis Processor.

The TMS5220 hardware and its software analogues (i.e. Arduino Talkie) alike may be driven by the output of TMS Express. The program accepts audio files in a variety of formats, applies linear predictive coding (LPC) to compress the data, and outputs either an ASCII bitstream or C header with the encoded data.

Compared to existing encoders, TMS Express has the following advantages:

Implements the original Levinson-Durbin recursion to determine LPC coefficients
Supports repeat frames for enhanced audio compression
Automatically downsamples and mixes audio files of any format
Performs batch encoding of multiple files

Installation

Install via Homebrew 🍺

Homebrew is the preferred installation method on macOS and Linux.

$ brew tap tornupnegatives/tmsexpress && brew install tms-express

Compile from Source

Dependencies

# Ubuntu
$ sudo apt install cmake libsndfile1-dev libsamplerate0-dev qt6-base-dev qt6-multimedia-dev

# macOS
$ brew install cmake libsamplerate libsndfile pkg-config qt

Compilation

$ cmake -B build && cd build
$ cmake --build . -j

Usage

The Encode Command

The encode command accepts audio file(s) and a variety of parameters which affect how they are processed, analyzed, and formatted for output. TMS Express automatically detects when the input path is a directory and performs a batch job.

$ tmsexpress encode [OPTIONS] input output

Explanation of Options

window: Speech data is separated into small windows/segments, each of which are analyzed individually. This is because small enough segments of speech data are roughly periodic and their behavior may be generalized. An ideal window width is between 22.5-25 ms
- Values above and below the recommendation will artificially speed up and slow down speech, respectively
highpass and lowpass: Speech data occupies a relatively small frequency band compared to what digital audio files are capable of representing. Filtering out unnecessary frequencies may lead to more accurate LPC analysis
- Lowering the highpass filter cutoff will improve the bass response of the audio
- Adjusting the lowpass cutoff may have minor effects of pitch estimation
alpha: While the pitch of speech is characterized by the lower frequency band, LPC algorithms which characterize the upper vocal tract benefit from an exaggeration of high frequency data. A pre-emphasis filter will exaggerate this part of the spectrum and lead to crisper, more accurate synthesis
format: ASCII format is ideal for testing and visualization of single files. The C and Arduino format produce C headers for use with TMS5220 emulations
no-stop-frame: An explicit stop frame signals to the TMS5220 that the Speak External command has finished executing
gain: Increases the gain of the synthesized signal by adjusting the index of the coding table element. Gain offsets greater than the max size of the coding table will hit the ceiling
max-voiced-gain: Specifies the maximum gain (dB) of the output signal for voiced frames (vowels)
max-unvoiced-gain: Specifies the maximum gain (dB) of the output signal for unvoiced frames (consonants)
- Ensuring that this value hovers around 0.8 * max-voiced-gain will result in the most accurate synthesis of consonant sounds
use-repeat-frames: Detect repeat frames to reduce the size of the bitstream
max-frq: Specifies the maximum representable pitch frequency of the output signal
min-frq: Specifies the minimum representable pitch frequency of the output signal

michaelkamprath/TMS-Express