Sound very quiet after applying de-emphasis

[KurtWoloch]

Under the "Output Options", you can opt to apply de-emphasis. However, if you do this, you end up with a very quiet waveform... usually much quieter than the source WAV file prior to emphasis. I think the volume should be boosted or normalized.

[cgreening]

Yeah, to be honest, I'm not sure about the de-emphasis - it's supposed to undo the emphasis, but I think what happens is:

Apply emphasis -> low frequencies suppressed, high frequencies increased.
Apple de-emphasis -> high frequencies supprussed.

So I wonder if we just end up with quieter audio with no benefit...

[KurtWoloch]

Well, the TI-99 speech synthesizer did seem to have some kind of de-emphasis going on since it doesn't sound as scratchy as if you just decode the LPC without de-emphasizing it. I think the option should be left in, but maybe it would be better to boost or normalize the audio.

[raphlinus]

The TI chip effectively undoes the preemphasis for voiced frames, as the chirp sequence is not too far off from an RC lowpass filter, which preemphasis inverts.

(The yellow line is the frequency response of the one-pole filter at 0.9735, while the blue line is the frequency response of TI_LATER_CHIRP from the mame sources)

Therefore there's no need to apply additional lowpass filtering. However, preemphasis should only be applied to voiced frames, as excitation for unvoiced frames has a flat white noise spectrum. Applying preemphasis to all frames leads to a highpass effect for unvoiced frames.

You should get even more accurate reproduction by using the inverse spectrum of the chirp, and my work-in-progress encoder does that.

[cgreening]

ok, just to make sure I understand the flow.

On the input waveform - we apply the emphasis to all samples.

On the output - we only want to remove the emphasis on voiced frames?

[raphlinus]

Here's the flow as I imagine it.

On encoding, first thing to do on each frame is classify voiced/unvoiced (note: BlueWizard doesn't do this, it classifies based on k1 from the encoding process).

Then, if voiced, apply a preemphasis filter which is ideally the inverse spectrum of the chirp pulse, which in turn may be truncated to the period for high pitches. However, a simple -0.9375 is pretty close and might be good enough.

On output, your goal is to match the TMS5220 closely. Your player and mine are giving pretty different results, yours sounds tinnier. I believe the main reason for this is that the chirp table in your player matches TI_0280_PATENT_CHIRP (from mame), while I'm using TI_LATER_CHIRP (which is what the TMS5220 uses, according to mame).

It's likely that actual synth modules had a lowpass filter after the raw chip output, probably to reduce the aliasing (bit-crushed) sounds from the 8k sampling rate. I wasn't able to find any definitive information on the recommended design of this filter; there's a schematic in Designing a TMS5220 Speech Synthesis Multimodule (video).

[KurtWoloch]

OK. Haven't had time to test the new version yet, but as we're talking about this topic, I found a website which may be useful for this purpose:
https://www.[unige.ch/medecine/nouspikel/ti99/speech.htm](https://www.unige.ch/medecine/nouspikel/ti99/speech.htm)

What strikes me about this write-up is the use of an interpolator for the parameters. That is, all parameters like pitch, energy and k values don't jump from value to value with each frame, but are interpolated so that throughout a frame, they move from the old values to the new values in 8 steps to reach the new value at the end of the frame. This would help smoothing out the speech output, including preventing jarring pitch jumps. I don't know if it interpolates all parameters or only certain ones though...
However, the page doesn't say anything about an additional low-pass filter, only that the output is connected to the sound chip via an 1 µF capacitor and a 330 Ohm resistor in series.

[raphlinus]

Yes, interpolation is a pretty important quality boost, and of course required for fidelity to the TI chips. It approximately linearly interpolates all parameters: energy, pitch, k, but there is also logic to inhibit interpolation when switching from unvoiced to voiced or vice versa, and in certain (chip-dependent) conditions involving silent frames. The mame source should be considered authoritative here, and I also have it implemented in my work-in-progress code.

speech_no_interp.wav
speech.wav

[KurtWoloch]

One more interesting source would be the TMS 5220 Voice Synthesis Processor Data Manual, which you can find here, among other places:
https://archive.org/details/bitsavers_tidataBooksisProcessorDataManualpreliminaryJun81_7901308/mode/2up

(I'm actually just sorting papers and stumbled upon a printed copy of it which I printed out at work from the PDF at a time when I didn't have Internet access at home yet, and I'm about to throw out this printed copy as well as some other printouts of manuals)

This manual also doesn't mention anything about de-emphasis going on after the decoding process.
Interestingly, the manual contains a calculation error (probably only the misspelling of a value)... it gives the interpolation interval at 3.25 instead of 3.125 milliseconds, but then goes on to calculate with 3.125.

[raphlinus]

I found an authoritative source for the lowpass filter: figure 1-7 of TSP50C0X/1X Family Speech Synthesizer Design Manual. That has a 1-pole (RC) lowpass filter with f_c at 3.6kHz, which makes a lot of sense as a value to reduce aliasing artifacts without seriously compromising speech bandwidth.

That document is also a goldmine when it comes to software implementation of the algorithm; it contains an annotated assembly code listing of code to synthesize speech from LPC bitstreams. That code shows an implementation of linear interpolation for the coefficients. This is approximated in the earlier chips by a shift-and-add operation per interpolation period, almost certainly to avoid the need for another multiplier. I think it's a good model for software which is trying to get the best quality speech from existing bitstreams using modest resources, as opposed to the MAME goal of fanatically correct emulation of old hardware.

That data manual is interesting but doesn't spill the secrets of how to do high quality encoding. The atariage forum introducing BlueWizard has some excellent background, including historical notes from Richard Wiggins of TI about their encoding pipeline.

[cgreening]

ok, I've added a low pass filter and I've added interpolation - it's not an exact match, but close enough.

This part I'm struggling with - we need to know if it's voiced or unvoiced to know if we should apply pre-emphasis. I've got a chicken and egg problem :)

On encoding, first thing to do on each frame is classify voiced/unvoiced (note: BlueWizard doesn't do this, it classifies based on k1 from the encoding process).

Then, if voiced, apply a preemphasis filter which is ideally the inverse spectrum of the chirp pulse, which in turn may be truncated to the period for high pitches. However, a simple -0.9375 is pretty close and might be good enough.

[raphlinus]

I've experimented with doing the encoding a second time, without preemphasis, for unvoiced frames. But my current code makes the voiced/unvoiced classification based on the confidence of pitch detection, before doing any LPC calculation.

Btw, I have a little more insight into the role of k1. I believe it represents the overall highpass/lowpass nature of the filter. The more negative, the more lowpass. So it's basically saying if there's a lot of high frequency content, it's unvoiced. That does seem like a reasonable heuristic to me. I have the feeling there's quite a bit of tuning that can be done on this, as it's one of the major factors affecting the smoothness or harshness of the encoded speech.