Incorrect conversion of uint8 WAV file in playback

[vladkorotnev]

Problem Description

Thanks for the nice project, it saved me in a pinch :-) (quickly upcycling an ESP32 with a fried LNA output into an audio toy for a friend's kid)

However I've encountered an issue with the resampling code: When playing a WAV file which is saved as 8 bits PCM via a FormatConverterStream, it is very noisy and distorted.

The reason seems to be because the samples are interpreted as signed 8 bit integers, which are not part of the WAV standard.

Device Description

ESP32 WROVER, but I think it's not device specific

Sketch

AnalogAudioStream out;
WAVDecoder decoder;

void init_wav() {
    AudioToolsLogger.begin(Serial, AudioToolsLogLevel::Error);
    auto config = out.defaultConfig();
    config.sample_rate = 44100; // <- for some reason ESP32 cannot start with any other sampling rate... I have this problem in my own projects too
    config.channels = 1;
    config.bits_per_sample = 16;
    out.begin(config);
}

void play_wav(const char * name) {
    File f = LittleFS.open(name);
    if(!f) return;

    EncodedAudioStream *s = new EncodedAudioStream(&f, &decoder);
    FormatConverterStream  *rss = new FormatConverterStream(*s);
    StreamCopy * copier = new StreamCopy(out, *rss);
    const AudioInfo from(16000, 1, 8);
    const AudioInfo to(44100, 1, 16);

    s->begin();
    rss->begin(from, to);
    copier->copyAll(1);

    f.close();
    delete rss;
    delete s;
    delete copier;
}

void setup() {
    if(!LittleFS.begin(false, "/disk", 128)) ESP_LOGE(LOG_TAG, "Mount fail");
    init_wav();
    play_wav("/test_8bit_16khz_mono.wav");
    vTaskDelete(NULL);
}

Other Steps to Reproduce

The following patch for AudioStreamsConverter.h fixes it, but breaks all other formats of course without proper integration:

378,379c379,386
<       NumberConverter::convertArray<TFrom, TTo>(
<           data_source, (TTo *)buffer.data(), samples, gain);
---
>       TTo* dest = (TTo *)buffer.data();
>       uint8_t *src = (uint8_t *)data;
>       for(int i = 0; i < samples; i++) {
>         int8_t sgn = ((int16_t)src[i]) - 127;
>         dest[i] = NumberConverter::clipT<int16_t>( ((int32_t)sgn * 256 - 128) * 3 );
>       }
>       // NumberConverter::convertArray<TFrom, TTo>(
>       //     data_source, (TTo *)buffer.data(), samples, gain);

What is your development environment

ESP32 + PlatformIO (espressif32 + arduino latest)

I have checked existing issues, discussions and online documentation

• [X] I confirm I have checked existing issues, discussions and online documentation

[pschatzmann]

This is correct: the FormatConverter only supports signed data types and the output of 8 bit via the ESP32 I2S is even messier because it expects an int16_t where the signed audio byte is left shifted by 1 byte.

You can chain another EncodedAudioStream with a DecoderL8 which supports both signed and unsiged bytes to convert it into a proper stream of int16_t

[vladkorotnev]

Oh, that's good to know. I have to admit this was on a whim and somehow didn't come up to this decoder via Google. So went with the good ol' peek and poke approach :-) Thanks for the pointer!

The esp was already decapitated and submerged into epoxy but it's good to know for later! Thanks so much!

P.S.

t expects an int16_t which is left shifted by 1 byte. The deeper I dive into this platform the more I am reassured it is cursed by design...