Man Goes Down - spectrogram improvements today!

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**AtariVox / SpeakJet emulation: Man Goes Down's speech, chip against library (0.4.2)**

These pictures compare the speech in Man Goes Down as two real AtariVoxes play it against the emulation as it is now (version 0.4.2, released tonight).

**What the pictures show.** Each one is a spectrogram: time across, frequency up (to 6 kHz), loudness as darkness. Each has four rows:

- the AtariVox+ ("unit A")
- the original AtariVox ("unit B")
- the library BEFORE: as it was at the start of tonight
- the library AFTER: version 0.4.2

Every row starts at its own first sound, so they line up by eye. The figure in each row's title is the phrase's length from its first sound to its last. The faint horizontal lines in the recordings are the recording setup, not the chip.

phrase_mgd_p1_before_after.png
- **phrase_mgd_p1_before_after.png:** phrase 1, "man goes down". Before, there's no gap between "man" and "goes", and "down" ends early. After, both line up with the AtariVoxes.

phrase_mgd_p1_goes_zoom_before_after.png
- **phrase_mgd_p1_goes_zoom_before_after.png:** the same phrase around "goes", 0.35-0.65 s and 300-3000 Hz. Look for the white gap before the G, which the chips have and BEFORE doesn't, and the two bands that follow it.

phrase_mgd_p2_before_after.png
- **phrase_mgd_p2_before_after.png:** phrase 2. It barely changes, which is the point: it was already right.

phrase_mgd_p4_before_after.png
- **phrase_mgd_p4_before_after.png:** phrase 4 (phrase 3 is the death sound, an effect with no words). Before, the hiss, the long vowel and the final "k" all come early. After, they line up.

**Where it started tonight, and where it is now:**

length (ms)    unit A   unit B   start of tonight   now
phrase 1        1495     1502         1388           1502
phrase 2        1695     1700         1695           1700
phrase 4         942      945          840            925

Phrase 1 was about 110 ms short, and its middle word sounded more like "gone" than "goes". Phrase 4 was about 100 ms short. Phrase 2 was already the right length, and the changes below had to keep it there while fixing the other two. Now the words and gaps of all three line up with both AtariVoxes to within about 20 ms.

**What it took.** Almost none of it was in the individual sounds, which already matched the chip on their own. It was in how the chip joins sounds, pauses and stretches them inside a phrase:

- **Speed below the default.** Phrase 1 plays at Speed 108. Below the default 114 the chip stretches speech the way it stretches the sound effects, more than the library did: x1.21 at Speed 108, not x1.125. That alone was most of phrase 1's missing time.

- **The G in "goes".** When a voiced stop like GO follows a nasal like NE, the chip goes almost silent for 20-30 ms before the burst, in every band, and the nasal's formants fall into the silence. The library ran straight through with its voice humming, so the G vanished. Then, before an OW, GO's release holds two bands near 1450 and 1950 Hz that split into the vowel. The zoomed picture shows both.

- **Joins between two sounds.** The library cut a fixed 30 ms off the first of two joined vowels. The pairs recorded in earlier probes say the chip takes longer to move between sounds whose formants are far apart: AW into IY is about 45 ms longer than the library had it, AW into UW slightly shorter. A stop running into a vowel follows a similar rule. There was also a plain bug: the rule for a stop before a vowel only checked the second sound, so GO lost 51 ms of its release.

- **Pauses.** Three of the pause codes are shorter on the chip than the manual says. P3, P5 and P6 run 30-50 ms under it, and P0 leaves about 20 ms, where the library had none.

- **Phrase 4.** It sends SE twice. A sound sent straight after itself plays as a repeat, and SE's repeat was too short (now 80 ms). The vowel before the hiss bends its formants up toward Z's, as both AtariVoxes show. And the slowed IYUW at its start runs about 50 ms longer on the chip than anywhere else it's been recorded.

Some of these rules rest on very few recordings: the vowel joins, the stop-into-vowel rule, GO before OW, and that phrase 4 IYUW. In the library's source they're marked as assumptions. VoxProbe 13, posted earlier tonight, tests each of them on both units, so they'll be confirmed or corrected when it's recorded. Two small things in phrase 4 are still wrong and visible: a dark blob inside the double SE, and a dark column where the hiss meets the UX.

One lesson along the way: the automatic scores I'd been using can't judge timing inside a phrase. Moving 5 ms of silence in a gap changed them more than any real fix did. So every change here was judged the old-fashioned way, by the length of each word and gap against both AtariVoxes, and by eye on these spectrograms.

The library is on GitHub as andrew-davie/SpeakJet (C, v0.4.2). The Go version is in the open pull request to JetSetIlly/SpeakJet. The technical manual in the C library, section 6.8, lists exactly what was fitted on Man Goes Down and how.


Andrew Davie

#1
The attached audio files play (after one beep) the original Man Goes Down sound, recorded from an actual Atari 2600/AtariVox.  Then (after two beeps) the library's version of the same.  These are awesomely close to my tired old ears.


alex_79