GROK LANGUAGE MODEL
CASE 01 · THE LAST SPEAKER
CA 0xfc27ab8c7e4fce08b2f437f31baf43fd5c580e9f
HOME
HOW TO CHECK
RULES
FIGHTS
CODE
X
TAPE POSITION 00000
a language with no alphabet, and one man left who speaks it
six agents listened to 1.2 hours of raw tape and pulled out
3,218 word candidates and 19 grammar rules - with no dictionary,
no translation and nobody alive who could tell them they got it wrong.
nothing is written without a source: every single number on this page opens the exact second of the recording it came from. the 41 rules the desk threw out are published too, with the reason each one died.
nothing is written without a source: every single number on this page opens the exact second of the recording it came from. the 41 rules the desk threw out are published too, with the reason each one died.
0.93 H OF SPEECH · 1,164 UTTERANCES
what this page proves, and what it does not
you can verify this yourself
- the tape is real and so are the positions. click any timestamp and the recording plays right here, at that second. cmd+click opens the full tape on the source.
- the chains really do repeat. a "word" here means an exact sequence of sounds that came back 14, 11, 9 times across 1,164 utterances. hear them and count.
- the desk refuses more than it accepts. 41 of 60 rules were thrown out, each with the reason, listed below.
- it is not pattern noise. 1,110 random chains of the same lengths were run through the same grammar: 0 of them land anywhere on this tape.
- the code is open. same pipeline, your own recording, one command.
what it is not
- nothing here is a translation. no dictionary was used and the desk does not know what any of it means. it never guesses a meaning.
- "word" and "rule" are mechanical terms. a word is a repeating chain of sounds, a rule is a chain that attaches to many different stems. a linguist would call this a starting point, not a description of the language.
- no native speaker has confirmed any of it. that is the whole problem of case 01: there is nobody left to ask.
- the serial on the account is dramatised. the desk, the tape and every number here are real. the story around them is a story.
how to check this yourself · takes about a minute
1
click any timestamp
every green position on this page is a link. it opens the recording at that exact second, so you hear the word instead of reading a claim about it.
2
read what got refused
the desk wrote 60 rules and signed 19. the other 41 are listed below with the reason each one failed the gate.
3
run it on your own tape
the code is open. desk.py run tape.wav, about a minute of cpu per hour of audio.
same pipeline, your recording.
tape: WIKITONGUES: Abdul Mateen speaking Torwali · Wikitongues, CC BY · every timestamp below opens that second of the recording
WORDS SAVED3,218
HEARD ONLY ONCE1,316
RULES SIGNED19
RULES REFUSED41
FIGHTS OPEN9
NOT HIS USUAL VOICE60
SOUNDS · CORE115 · 45
SPEECH ON TAPE0.93 h
SPEAKERS LEFT1
NEVER RECORDED1,105
DIE WITH HIM2,041
the wall · chains that keep coming back
rules signed by CHIEF · a chain that attaches to many different neighbours
| 1 | suffix | ʌ ɾ ɒ | 11× on 10 stems | 00:25:28 00:25:32 00:30:59 |
| 2 | prefix | b u l | 10× on 10 stems | 00:02:37 00:13:18 00:13:49 |
| 3 | suffix | ɒ x ɒ | 10× on 9 stems | 00:02:18 00:04:39 00:11:24 |
| 4 | prefix | b æ n | 9× on 9 stems | 00:03:50 00:19:55 00:22:35 |
| 5 | suffix | i tɕ i | 9× on 8 stems | 00:01:16 00:15:54 00:25:48 |
| 6 | suffix | b a n | 8× on 8 stems | 00:13:27 00:24:46 00:25:09 |
| 7 | prefix | b a ɾ | 8× on 8 stems | 00:27:15 00:29:23 00:43:13 |
| 8 | prefix | ʔ ɒ l | 8× on 7 stems | 00:00:24 00:03:58 00:05:32 |
| 9 | prefix | ʔ ɒ tɕ | 8× on 7 stems | 00:00:51 00:07:20 00:07:46 |
| 10 | prefix | b u ɾ | 8× on 7 stems | 00:04:16 00:05:04 00:07:29 |
| 11 | prefix | s ɒ n | 8× on 7 stems | 00:09:33 00:10:46 00:20:37 |
| 12 | suffix | a l a | 8× on 7 stems | 00:16:41 00:28:43 00:35:27 |
| 13 | suffix | l j i | 7× on 7 stems | 00:01:49 00:29:07 00:41:12 |
| 14 | suffix | i ɕ i | 7× on 7 stems | 00:02:41 00:08:10 00:09:16 |
| 15 | prefix | a tʂ e | 7× on 7 stems | 00:20:59 00:23:29 00:28:32 |
| 16 | prefix | ʔ ɒ n | 7× on 6 stems | 00:00:17 00:03:58 00:10:52 |
| 17 | prefix | b o ɾ | 6× on 6 stems | 00:07:20 00:21:02 00:46:52 |
| 18 | suffix | b ɒ ɾ ɒ | 6× on 5 stems | 00:03:41 00:08:06 00:09:41 |
| 19 | prefix | tɕ ɒ b | 6× on 5 stems | 00:10:57 00:15:45 00:31:06 |
fights · two answers, nobody alive to settle it
| b u l | vs | b u ɾ | one phone apart, both come back: two answers |
| i tɕ i | vs | i ɕ i | one phone apart, both come back: two answers |
| b a ɾ | vs | b u ɾ | one phone apart, both come back: two answers |
| b a ɾ | vs | b o ɾ | one phone apart, both come back: two answers |
| ʔ ɒ l | vs | ʔ ɒ tɕ | one phone apart, both come back: two answers |
| ʔ ɒ l | vs | ʔ ɒ n | one phone apart, both come back: two answers |
| ʔ ɒ tɕ | vs | ʔ ɒ n | one phone apart, both come back: two answers |
| b u ɾ | vs | b o ɾ | one phone apart, both come back: two answers |
| s ɒ n | vs | ʔ ɒ n | one phone apart, both come back: two answers |
the gate · refused (41)
- ʔ ɒ tʂ · heard 5 times, the gate wants 6
- tɕ ɒ l · heard 5 times, the gate wants 6
- o b ɒ · heard 5 times, the gate wants 6
- n ɒ ɴ · heard 5 times, the gate wants 6
- o b u · heard 5 times, the gate wants 6
- ð e ɾ · heard 5 times, the gate wants 6
- i j b · heard 5 times, the gate wants 6
- ʌ b ɒ · heard 5 times, the gate wants 6
- b ɔ ɾ · heard 5 times, the gate wants 6
- ɨ b ɒ · heard 5 times, the gate wants 6
- … and 31 more in report.md
his voice · where the register shifts (median 152.9 Hz, 10.27 phones/s)
| 00:38:45 | 332.2 Hz | z +7.4 | 2.86 phones/s |
| 00:52:53 | 299.1 Hz | z +6.1 | 3.75 phones/s |
| 00:49:40 | 290.5 Hz | z +5.7 | 2.5 phones/s |
| 01:04:48 | 275.4 Hz | z +5.1 | 2.73 phones/s |
| 01:06:09 | 271.9 Hz | z +4.9 | 6.25 phones/s |
| 01:11:42 | 270.6 Hz | z +4.9 | 5.0 phones/s |
| 00:07:15 | 260.2 Hz | z +4.5 | 8.0 phones/s |
| 01:08:06 | 253.5 Hz | z +4.2 | 7.35 phones/s |
| 00:56:22 | 246.8 Hz | z +3.9 | 11.82 phones/s |
| 00:21:32 | 242.4 Hz | z +3.7 | 14.5 phones/s |
| 00:18:16 | 237.9 Hz | z +3.5 | 1.0 phones/s |
| 00:21:37 | 234.5 Hz | z +3.4 | 8.75 phones/s |
vowels nobody wrote down
ɒ3353
e1912
a1851
o1089
ʌ1041
uə994
i967
ɪ611
æ543
ɨ299
ɔ291
u268
case log
16:29:40 CHIEF case open · torwali.wav · 1.2 h of tape · one speaker · no dictionary, no translation
16:29:52 SCOUT cut the tape into 1164 utterances · 55.9 min of speech in 71.9 min of tape
16:29:53 PHON utterance 0 · 00:00:01 · 57 phones · inventory 26
16:29:55 PHON utterance 50 · 00:03:18 · 43 phones · inventory 88
16:29:57 PHON utterance 100 · 00:06:05 · 27 phones · inventory 96
16:29:59 PHON utterance 150 · 00:08:51 · 64 phones · inventory 101
16:30:00 PHON utterance 200 · 00:11:35 · 8 phones · inventory 104
16:30:02 PHON utterance 250 · 00:14:52 · 40 phones · inventory 106
16:30:04 PHON utterance 300 · 00:17:55 · 16 phones · inventory 108
16:30:06 PHON utterance 350 · 00:20:32 · 16 phones · inventory 109
16:30:07 PHON utterance 400 · 00:23:03 · 21 phones · inventory 110
16:30:09 PHON utterance 450 · 00:25:43 · 8 phones · inventory 112
16:30:11 PHON utterance 500 · 00:29:17 · 6 phones · inventory 113
16:30:13 PHON utterance 550 · 00:32:28 · 7 phones · inventory 113
16:30:15 PHON utterance 600 · 00:35:37 · 45 phones · inventory 113
16:30:17 PHON utterance 650 · 00:39:45 · 48 phones · inventory 113
16:30:19 PHON utterance 700 · 00:42:50 · 55 phones · inventory 113
16:30:21 PHON utterance 750 · 00:46:26 · 58 phones · inventory 113
16:30:23 PHON utterance 800 · 00:49:13 · 22 phones · inventory 113
16:30:25 PHON utterance 850 · 00:52:34 · 16 phones · inventory 113
16:30:27 PHON utterance 900 · 00:55:19 · 10 phones · inventory 114
16:30:29 PHON utterance 950 · 00:58:38 · 41 phones · inventory 114
16:30:31 PHON utterance 1000 · 01:01:54 · 9 phones · inventory 115
16:30:33 PHON utterance 1050 · 01:04:55 · 9 phones · inventory 115
16:30:35 PHON utterance 1100 · 01:07:43 · 57 phones · inventory 115
16:30:37 PHON utterance 1150 · 01:11:02 · 22 phones · inventory 115
16:30:37 PHON sound inventory from scratch: 115 phones · 23 vowels · 92 consonants
16:30:37 LEX 3218 word candidates (chains that come back ≥3× in ≥3 utterances) · 7750 tokens · 72% of the tape covered · 1316 stretches heard exactly once, so they could mean anything
16:30:38 GRAM 60 rules written (28 endings, 32 prefixes) · 38 of them fight each other
16:30:39 VOICE his voice: median pitch 153 Hz, 10.3 phones/s · 60 utterances where the register shifts (pitch up, tempo up: not his usual voice)
16:30:39 CHIEF signed 19 rules · refused 41 · 9 fights stay open, nobody alive to settle them
16:30:39 CHIEF refused rule 20 [ʔ ɒ tʂ] · heard 5 times, the gate wants 6
16:30:39 CHIEF refused rule 21 [tɕ ɒ l] · heard 5 times, the gate wants 6
16:30:39 CHIEF refused rule 22 [o b ɒ] · heard 5 times, the gate wants 6
16:30:39 CHIEF refused rule 23 [n ɒ ɴ] · heard 5 times, the gate wants 6
16:30:39 CHIEF refused rule 24 [o b u] · heard 5 times, the gate wants 6
16:30:39 CHIEF refused rule 25 [ð e ɾ] · heard 5 times, the gate wants 6
16:30:39 CHIEF ... and 35 more refusals, all in report.md with the reason
16:30:39 SCRIBE filed run.json, report.md, log.txt in /Users/wooooork/projects/tools/last-speaker/runs/torwali
16:30:39 CHIEF done in 1.0 min · WORDS SAVED 3218 · RULES 19 · FIGHTS 9 · SHIFTS 60 · SPEAKERS LEFT 1
forms this grammar allows that nobody ever said
run the signed rules the other way: if one stem takes an ending and another stem takes a different one, both crossings are forms the language permits. every one of them was then searched across all 1,164 utterances as an exact phone chain. 1,110 forms · 5 turned out to be on the tape after all · 1,105 are on no second of it (first 14 below).
control: 0 of 1,110 RANDOM chains of the same lengths land anywhere on this tape. without that line the hits above would mean nothing.
control: 0 of 1,110 RANDOM chains of the same lengths land anywhere on this tape. without that line the hits above would mean nothing.
| j ʌ m ʌ ɾ ɒ | j ʌ m | + rule 1 | ʌ ɾ ɒ | stem 2× · rule 11× |
| j ʌ m ɒ x ɒ | j ʌ m | + rule 3 | ɒ x ɒ | stem 2× · rule 10× |
| j ʌ m i tɕ i | j ʌ m | + rule 5 | i tɕ i | stem 2× · rule 9× |
| j ʌ m b a n | j ʌ m | + rule 6 | b a n | stem 2× · rule 8× |
| j ʌ m l j i | j ʌ m | + rule 13 | l j i | stem 2× · rule 7× |
| j ʌ m i ɕ i | j ʌ m | + rule 14 | i ɕ i | stem 2× · rule 7× |
| j ʌ m b ɒ ɾ ɒ | j ʌ m | + rule 18 | b ɒ ɾ ɒ | stem 2× · rule 6× |
| ʂ ɨ b ɒ x ɒ | ʂ ɨ b | + rule 3 | ɒ x ɒ | stem 2× · rule 10× |
| ʂ ɨ b i tɕ i | ʂ ɨ b | + rule 5 | i tɕ i | stem 2× · rule 9× |
| ʂ ɨ b b a n | ʂ ɨ b | + rule 6 | b a n | stem 2× · rule 8× |
| ʂ ɨ b a l a | ʂ ɨ b | + rule 12 | a l a | stem 2× · rule 8× |
| ʂ ɨ b l j i | ʂ ɨ b | + rule 13 | l j i | stem 2× · rule 7× |
| ʂ ɨ b i ɕ i | ʂ ɨ b | + rule 14 | i ɕ i | stem 2× · rule 7× |
| ʂ ɨ b b ɒ ɾ ɒ | ʂ ɨ b | + rule 18 | b ɒ ɾ ɒ | stem 2× · rule 6× |
the queue · what goes when he goes
a word the rules can rebuild after he is gone is not the urgent one. a chain heard once or twice that no rule produces is the one nobody recovers. 2,041 of 3,218 word candidates are in that second group - that is the order to work through, not the frequency list.
| t uə e | 1× in 1 utterances | nothing can rebuild it |
| b a l | 1× in 1 utterances | nothing can rebuild it |
| i ɕ a | 1× in 1 utterances | nothing can rebuild it |
| e ʂ ɒ | 1× in 1 utterances | nothing can rebuild it |
| ɪ b ʌ | 1× in 1 utterances | nothing can rebuild it |
| ʌ n d i | 1× in 1 utterances | nothing can rebuild it |
| n t i ɕ i | 1× in 1 utterances | nothing can rebuild it |
| i x a | 1× in 1 utterances | nothing can rebuild it |
| ɒ l b | 1× in 1 utterances | nothing can rebuild it |
| tʂ ʌ ɒ | 1× in 1 utterances | nothing can rebuild it |
| ʁ ɒ tɕ | 1× in 1 utterances | nothing can rebuild it |
| e m j | 1× in 1 utterances | nothing can rebuild it |
| i ʂ ʌ | 1× in 1 utterances | nothing can rebuild it |
| i b u l | 1× in 1 utterances | nothing can rebuild it |
does the voice fade across the tape
tempo, speech share and utterance length in 10-minute windows. trend per window: rate -0.6%, speech_share +0.4%, utt_seconds +2.6%. a flat line is a result, not a missing measurement - the serial's voice budget needs sessions months apart, and one recording cannot give that.
| window | utterances | speech share | phones/s | pitch | median utterance |
| 00:00-10:00 | 170 | 0.773 | 10.0 | 156.8 | 2.1 s |
| 10:00-20:00 | 172 | 0.771 | 10.87 | 149.7 | 2.2 s |
| 20:00-30:00 | 171 | 0.778 | 11.43 | 153.1 | 2.0 s |
| 30:00-40:00 | 140 | 0.76 | 9.16 | 155.4 | 2.65 s |
| 40:00-50:00 | 156 | 0.799 | 10.0 | 153.4 | 2.3 s |
| 50:00-60:00 | 160 | 0.792 | 10.33 | 152.7 | 2.3 s |
| 60:00-70:00 | 167 | 0.764 | 10.0 | 152.1 | 2.2 s |
| 70:00-71:00 | 28 | 0.8 | 10.28 | 150.9 | 2.75 s |
what this is not
no dictionary of the language was used and nothing here is a translation. a "word" is a chain of phones that repeats, a "rule" is a chain that keeps attaching to different neighbours, a "fight" is two of them one phone apart. the register detector hears pitch and tempo, not meaning. a linguist would call this a starting point. that is what it is, and every line carries the second of tape it came from.
CONTRACT
0xfc27ab8c7e4fce08b2f437f31baf43fd5c580e9fTHE DESK DOES NOT TRADE · IT READS TAPE
TAPE
OPEN THE FULL TAPE ↗
other tapes: torwali · piedmontese
· code
· the serial
· desk.py run tape.wav · generated from runs/torwali/run.json