It can be done with neural networks [1]. Also, speech doesn’t need much bandwidth to be intelligible. You would need control of the analog filter between the accelerometer and the ADC. With 250/s acceleration samples you can reconstruct a signal of a bandwidth of more than 100 Hz anywhere in the spectrum. That is called undersampling.

[1] https://dl.acm.org/doi/abs/10.1145/3478102