Example 'micro_speech' hung when I2S config bits_per_sample changed to 24 from 16

[misterb0407]

Since I am using ESP-EYE which uses MSM261S4030H0R microphone, and according to datasheets it has 24-bit depth in a 32 bit wordk. So in i2s_init() I change the :

i2s_config_t i2s_config = {
    ...
    .bits_per_sample = (i2s_bits_per_sample_t)24, // was 16
}

But when loading it to the board, the console remain like this forever, regardless we speak over phone or not:

I (336) cpu_start: Starting scheduler on PRO CPU.
I (0) cpu_start: Starting scheduler on APP CPU.
I (430) I2S: DMA Malloc info, datalen=blocksize=1200, dma_buf_count=3
I (430) I2S: I2S1, MCLK output by GPIO0
I (490) TF_LITE_AUDIO_PROVIDER: Audio Recording started

By right we should expect print out to reflect "yes" or "no" being detected.

[vikramdattu]

@misterb0407 from the data sheet: The serial data is in slave mode I2S format, which has 24‐bit depth in a 32 bit word. That means, you need to set bits_per_sample to 32 bits and ignore MSB 8 bits. I could find an example retrieving data in this format here.

But since, the feature extractor from the example expects the data in 16 bits, is this something you want to do?

[misterb0407]

Hi @vikramdattu , my main objective is to get this micro_example works on my ESP-EYE board with 'decent' accuracy, Currently it is just so poor that I don't believe it is because of the model alone.

Anyway thanks for the link, I tried similar setting, but no improvement observed.