Terms used in the sotto code, the design notes in docs/, and this course, in plain words. Names in code font are how the term appears in the code. The chapter where each idea is explained is in brackets.
Saying the words. Where a word is hard to say, a sound-it-out spelling follows it, like fricative (say FRIK-uh-tiv). The part in CAPITALS is the one you stress. “Said as letters” means you spell it out, like “D-C-T”.
A (amplitudes): A in the model. The loudness of each harmonic in one 10 ms frame. A[1] is the first harmonic, A[L] the last. [2]
Analysis: the encoder’s first job: measuring pitch, voicing and harmonic amplitudes from the microphone signal. sotto uses Codec 2’s analysis, with its own pitch and voicing tracker on top (v3.5 uses analyse(x, "tracked3") in sim/vocoder_lab.py; the tracker is in sim/sinusoid.py). [2]
AMBE / AMBE+2 (say AM-bee, AM-bee plus two): the closed, licensed vocoder used by Yaesu System Fusion and other digital radios. sotto aims to match its quality with open code.
Aspiration (say as-puh-RAY-shun): the breathy puff of hiss after a “t”, “p” or “k” burst, before the vowel’s buzz starts. [8]
Band voicing: mixed voicing done simply: harmonics below a cutoff frequency are voiced and those above it get random phases. Tested for v1 with no PESQ gain, then brought back in v3.2 after listeners heard “five-anuh”. sotto’s cutoff is 1,000 Hz (MIX_HZ, mixed_cutoff()). See mixed voicing. [2, 5]
Bit: a single 0 or 1. With n bits you can send one of 2ⁿ messages. [1, 4]
Bit allocation: deciding how many bits each number in a frame gets. sotto’s old scalar shape quantizer used greedy bit allocation: it handed out bits one at a time, each to whichever number it helped most. [4]
Blind A/B test: a listening test where the listener hears two versions of a clip, labeled only A and B, and says which is better or that they’re the same. They don’t know which codec is which. See listening round. [7]
Bit rate: bits per second. sotto’s vocoder runs at 2,400 bits per second; the radio carries 9,600. [1, 6]
Blend (hiss blend): in the v3.8 decoder, one hiss layer above 4,000 Hz made from the louder of the consonant hiss and the vowel layer at each frequency, so the hiss doesn’t pop on and off when frames flicker between unvoiced and mixed (blend=True). [8]
Burst (stop burst): the short pop of air when a “t”, “p” or “k” is released, about 10 ms long. The v3.8 burst rule sends a frame where 4,000 to 8,000 Hz jumps by more than 12 dB in 10 ms as U U, so it gets a hiss index. See stop consonant. [8]
CC BY: a Creative Commons license that lets anyone reuse a work if they credit the authors. sotto’s codebook is trained only on CC BY speech, so it can be shipped. [4, 7]
C4FM (say see-four-eff-em): “4-level FM”. The radio signal sotto sends: the frequency hops between four values (±600 and ±1,800 Hz) 4,800 times a second, carrying 2 bits per hop. [6]
C5517: the TI digital signal processor (DSP) chip inside the Yaesu FTM-300D radio, where sotto is meant to run one day. It has 320 KB of memory, which is why v3.6 shrank the codebook to 29 KB. [4]
c2sim, c2enc, c2dec (say see-two-sim, see-two-enk, see-two-dek): Codec 2’s command-line programs. c2sim runs the analysis and synthesis separately, so experiments can edit the model in between (--modelout, --modelin). c2enc and c2dec are the full Codec 2 encoder and decoder. [7]
Codebook: the list of allowed patterns in a vector quantizer, like a numbered color chart. The encoder and decoder have identical copies; the encoder sends the number of the closest entry instead of the pattern itself. In v3.5 the three stages have 4,096, 2,048 and 2,048 entries of 19 numbers each (152 KB at one byte per number). v3.6 shrank it to four stages of 512, 512, 256 and 256 entries (29 KB) so it fits in the radio’s DSP chip and searches about four times faster, with no audible loss. [4]
CODEC_DELAY: the fixed 160-sample (20 ms) delay used to line up every decoded clip with the original before scoring (sim/audio_race.py). It replaced a waveform-matching search that picked wrong delays for vocoders with made-up phases. [7]
Codec (say KOH-dek): short for coder-decoder. Any system that packs data into bits and unpacks it again. A vocoder is a codec for voice.
Codec 2: an older open-source vocoder by David Rowe. sotto borrows its analysis (including its pitch and voicing estimates, which sotto’s tracker starts from) and its synthesizer, and keeps a copy of its source code in codec/codec2/ (under its own LGPL license). [1]
Concealment: hiding a lost frame from the listener. Since v3.4 it’s done in sotto’s decoder: hold the last good frame for 40 ms, fade it by 6 dB every 20 ms, and go silent after 80 ms (SottoV34.decode(); the v3.5 settings are hold=2, slope=6, max_lost=4). Codec 2 instead repeats the last good frame up to 3 times at full level, then mutes. [6]
Convolutional code: an error-correcting code in which every output bit mixes the last several input bits (7 for sotto: “K = 7”). Decoded with the Viterbi algorithm. [6]
CRC (CRC-4, CRC-8, CRC-16): “cyclic redundancy check”. A short fingerprint (4, 8 or 16 bits) computed from a block of data, which lets the receiver tell whether the block arrived intact. It detects errors but doesn’t fix them. The speaker side channel uses a 4-bit one. [6, 8]
dB (decibel) (dB is said as letters; decibel is DESS-uh-bel): a scale for loudness or power that works by ratios. +10 dB is 10 times the power; +6 dB is about 4 times; +3 dB is about 2 times. [3, 4]
dB-Hz (C/N₀) (say dee-bee-HURTS; C/N₀ is “C over N zero”): how strong a radio signal is compared with the noise, measured per hertz of bandwidth. In the test conditions, 52 dB-Hz is a weak signal and 56 dB-Hz a somewhat better one. [6]
DCT (discrete cosine transform) (said as letters; cosine is KOH-sine): rewriting a list of numbers as the sizes of cosine waves added together. sotto applies it to the 20-point envelope shape, so most of the shape ends up in the first few numbers. JPEG uses the same idea. DCT in the code is the 20 × 20 table. [3]
Decimate / interpolate (say DESS-uh-mate, in-TUR-puh-late): decimating means keeping only every second frame (sotto sends one analysis per 20 ms out of two). Interpolating means filling the missing frames in with values halfway between their neighbors. decimate_interp(), interp_wo(). [4]
Decoder: the receiving half of a codec. sotto’s decoder turns 48 bits into 20 ms of sound. SottoV35.decode() (most steps are inherited from SottoV0.decode()). [5]
Discriminator (say dis-KRIM-uh-nay-ter): the part of an FM receiver that turns frequency back into a voltage. The softmodem reads its output. [6]
DN / VW: two voice layouts in System Fusion. DN (“digital narrow”, also called V/D mode 2) uses part of each frame’s payload for voice; VW (“voice wide”, also called voice full-rate) uses all 720 payload bits. sotto uses the VW slot, protecting its voice much more strongly. [6]
Doppler (fd) (say DOP-ler; fd is said as letters): how fast the fading changes, which depends on the radio frequency and how fast you’re moving. 446 MHz at 60 km/h gives about 25 Hz; 146 MHz at 30 km/h about 4 Hz. [6]
Encoder: the sending half of a codec. SottoV35.encode(). [1]
EARS: a collection of studio speech recordings. sotto uses it only for testing, because its license doesn’t allow shipping a codebook trained on it. [7]
Edge (voicing): one of the four voicing codes: one 10 ms half of the frame is unvoiced and the other mixed. Whether it’s a start (U then M) or an end (M then U) is worked out from the frame before. [2, 4]
Encoder race: an experiment where encoders (Codec 2 2400 and the sotto versions) send through the identical radio chain with the same fading, averaged over three fading patterns, and are scored the same way. sim/encoder_race.py, results in docs/results/encoder_race.json. [6, 7]
Energy: how loud a frame is: the sum of the harmonics’ powers, in dB. Sent with 5 bits from 10 to 80 dB. Since v3, code 0 means true silence (see silence). Since v3.6 it measures only the loudness inside the envelope’s range (200 to 3,700 Hz), so a hum below 200 Hz isn’t spread over the whole band. [4]
Envelope (spectral envelope) (say EN-vuh-lope; spectral is SPEK-trul): the smooth outline across the tops of the harmonics. It describes the shape of the throat and mouth, and carries most of the words. [3]
Erasure (say ih-RAY-shur): a bit or frame the receiver knows is missing or unreliable, as opposed to one that is silently wrong. [6]
ESTOI (say EE-stoy): a computer score that predicts intelligibility (how many words a listener would get), from 0 to 1. [7]
FEC (forward error correction): extra bits added before sending so the receiver can fix errors itself, without asking for a resend. sotto’s FEC is a convolutional code. [6]
FICH (said as letters): “frame information channel”. The 200-bit header in every System Fusion frame that says what kind of frame it is. sotto marks its frames with one spare bit in the FICH (a provisional choice). [6]
Flywheel: the receiver’s habit of expecting the next frame exactly 100 ms after the last, so it keeps its place through a fade. [6]
Formant (say FOR-munt; plural FOR-munts): a loud bump in the envelope, caused by a resonance of the throat and mouth. Different vowels have formants in different places. [3]
Formant postfilter: sotto’s decoder step that stretches the envelope’s peaks and valleys by 1.35 (in dB) to undo blurring from quantization (1.5 before v3.5). postfilter_rk(). Not the same as Codec 2’s postfilter. [5]
Frame: a chunk of data handled as one unit. A sotto vocoder frame is 20 ms and 48 bits. A System Fusion radio frame is 100 ms and 960 bits, carrying 5 vocoder frames. The analysis also works in 10 ms frames, sometimes called subframes. [1, 6]
Frequency (say FREE-kwun-see): how many times per second something repeats, in hertz (Hz, said hurts). [2]
Fricative (say FRIK-uh-tiv): a consonant made by pushing air through a narrow gap, so it hisses, like “f”, “s”, “sh” and “th”. The name comes from friction: the air rubs past the lips, teeth or tongue. Plain fricatives are unvoiced (all hiss). Ones with vocal-cord buzz as well, like “v” and “z”, are voiced fricatives. “s” and “sh” are the loudest and hissiest; in studio recordings most of an “s” is between 4,000 and 8,000 Hz, above the 4,000 Hz that sotto’s 8,000 samples per second can hold. That’s why sotto’s “s” sounds duller than the original. [1, 2]
Glorp: listeners’ word for the gargle heard when the woman’s pitch went above Codec 2’s 400 Hz limit. It led to sotto’s 500 Hz limit (v2). [2, 7]
Golay code (say goh-LAY): a classic error-correcting code that protects the FICH header. [6]
Greedy: an approach that makes the best-looking choice at each step without planning ahead. See bit allocation. [4]
Harmonics (say har-MON-iks): sine waves at whole-number multiples of the pitch. A voiced sound is the sum of its harmonics. [2]
Hiss band (high band): sotto v3.8’s 4,000 to 8,000 Hz layer of shaped noise, added on top of the normal 0 to 4,000 Hz output. It needs 16,000 samples per second in and out, but no extra bits (SottoV38 in sim/hiss_band.py, c/sotto_hiss.c). [8]
Hiss index: the 6 bits in a U U frame’s pitch field that describe the hiss band: one of 64 patterns of corrections to the decoder’s free guess of four band levels (4 to 5, 5 to 6, 6 to 7 and 7 to 8 kHz). [8]
Held out: test data kept out of training, so the test is fair. sotto’s codebook is tested on 16 held-out voices. [4, 7]
Hold and fade: the two halves of sotto’s concealment: hold repeats the last good frame for a short time, fade then lowers it step by step. [6]
Interleaving (say IN-ter-leev-ing): see spreading.
k-means (say KAY-means): a training method for finding a good codebook. Start with random entries, put each training example with its nearest entry, move every entry to the average of its group, and repeat until nothing changes. sotto trains each VQ stage this way, then re-tunes the stages together (ShapeMSVQ). [4]
L: the number of harmonics below 4,000 Hz: L = ⌊π ÷ Wo⌋. [2]
LibriSpeech: a large collection of audiobook recordings (CC BY 4.0). Part of the training speech for sotto’s codebook. [4]
Listening round: one batch of blind A/B tests played to listeners (members of a radio club). Each round includes the original clip next to every pair, and a null pair. [7]
LLR (log-likelihood ratio) (said as letters): a soft bit: a number saying how sure the receiver is that a bit is 0 (positive) or 1 (negative). Bigger means more sure. [6]
Lookahead: how far into the future an encoder must see before deciding about the present. sotto’s tracker can search the whole recording, but tests showed 80 ms of look-ahead loses only 0.003 PESQ, so the real-time version will use 80 ms (lag in track_pitch_voicing()). v3.8’s voicing rules need more: its encoder looks 10 frames (about 320 ms in all) ahead (SOTTO_RULE_LAG in the C port). See Viterbi algorithm. [2, 7, 8]
LPC / LPC-10: linear prediction. Codec 2 2400’s way of describing the envelope, with 10 numbers that define a smooth curve. The quantization ladder showed it’s Codec 2’s biggest quality loss for male voices. [3, 7]
LSP: “line spectral pairs”. Another way of writing the 10 LPC numbers that is easier to round off safely. Codec 2 2400 quantizes LSPs. [7]
M-best search: a shortcut for searching a multistage VQ. Instead of keeping only the single best choice after each stage, keep the 8 best so far and pick the overall winner at the end. It finds nearly the best combination without trying all 17 billion. [4]
Mel scale (rhymes with “bell”): a frequency scale that matches human hearing, finer at low frequencies and coarser at high ones. sotto’s 20 envelope points are evenly spaced on it, from 200 to 3,700 Hz. [3]
Memoryless: each frame can be decoded on its own, without the frames before it. sotto is memoryless so that one lost frame can’t damage the next ones. (The voicing edge code peeks at the frame before, but only to choose between two orderings.) [4]
Mixed voicing (mixed excitation, MBE, MELP) (MBE is said as letters; MELP rhymes with “help”): letting a sound be partly buzz and partly hiss, like “z” or “v”. Since v3.2, sotto has three kinds of frame: U (unvoiced, all noise), V (voiced, all harmonics) and M (mixed: harmonics below 1,000 Hz, noise above). A frame becomes M when it’s voiced but its harmonics fit poorly below 1,000 Hz (voicing_classes()). MBE and MELP split voicing into more bands. [2, 5]
Multistage VQ (MSVQ): a vector quantizer done in steps. Stage 1 picks the closest codebook entry; stage 2 picks an entry that best corrects what stage 1 got wrong; stage 3 corrects what’s left. sotto’s shape quantizer uses 3 stages of 12, 11 and 11 bits in v3.5 and 4 stages of 9, 9, 8 and 8 bits in v3.6, 34 bits either way (ShapeMSVQ). [4]
Model: the vocoder’s description of one 10 ms frame: Wo, L, A, phi and voiced. Stored in the MODEL record type in sim/vocoder_lab.py. [2]
newamp1 (say new-amp-one): an experimental Codec 2 mode that introduced the rate-K mel envelope and the formant postfilter that sotto builds on. [3]
Narrowband / wideband: speech at 8,000 samples per second (up to 4,000 Hz, like a phone call) or at 16,000 (up to 8,000 Hz). sotto up to v3.7 is narrowband; v3.8 plays wideband. PESQ comes in both kinds (PESQ-NB and PESQ-WB). [1, 8]
NLP: “non-linear pitch”. Codec 2’s pitch tracker. sotto v0 used it as is; later versions start from its answers and correct octave errors. [2]
Null pair: a listening-test pair where A and B are the same clip. A careful listener calls it a tie; a listener who picks a winner is probably guessing elsewhere too. [7]
Nyquist frequency (say NY-kwist): half the sample rate: the highest frequency a recording can hold. 4,000 Hz at 8,000 samples per second, 8,000 Hz at 16,000. [1, 8]
Octave error: a pitch tracker reporting double or half the true pitch. [2]
Odd/even check (sub-octave check): a test added to the tracker in v3.3. If a pitch candidate’s odd harmonics (3rd, 5th, 7th…) are much weaker than its even ones, it’s probably half the true pitch, so it gets a penalty. odd_even_db() in sim/sinusoid.py. [2]
Oracle: in an experiment, a helper that uses the true, perfect value instead of an estimate or a rounded value. Comparing with and without an oracle shows how much that part costs. [7]
Oracle ladder: see quantization ladder.
Oracle lab: sim/vocoder_lab.py: tools for editing Codec 2’s model in Python and listening to the result. [7]
Overlap-add (say OH-ver-lap add): building a long sound from short pieces that overlap, fading each one out as the next fades in. For noise it loses 1.76 dB of power, because random phases add by power, not like waves. [5, 8]
Payload (say PAY-lode): the part of a frame that carries the actual data (as opposed to sync and headers). In System Fusion, 720 bits per 100 ms. [6]
PCA (principal component analysis) (said as letters; principal is PRIN-suh-pul): a way of finding the few patterns that explain most of the differences in a big table of numbers. In v3.8, three patterns describe 92% of how speakers’ high bands differ, so a speaker’s shape is just 3 numbers. [8]
PESQ (say PESK): “perceptual evaluation of speech quality”. A computer score that predicts how listeners would rate quality, from about 1 (bad) to 4.5 (perfect). [7]
Phase (phi) (phi is said FYE, like “pie”): phi in the model. Where in its cycle a sine wave starts. sotto doesn’t send phases; the decoder invents them. [5]
phase0 (synthetic phase) (say phase-zero): Codec 2’s way of inventing phases: imagine each vocal-cord pulse as a click played through the envelope filter. Unvoiced frames get random phases. [5]
Pitch (F0): how fast the vocal cords buzz, in Hz. About 85–155 Hz for men and 165–255 Hz for women. [2]
Pitch oracle: an oracle test for pitch: the decoder is given the true (or carefully checked) pitch instead of the rounded or estimated one, to measure how much pitch errors cost. [7]
Pitch tracker (pitch estimator): the part of the encoder that measures pitch. Errors where it reports double or half the true pitch are called octave errors. See sotto tracker. [2]
Pitch-field leak: a v3.8 bug. When the decoder filled in a voiced 10 ms piece after an unvoiced half, it borrowed the next frame’s pitch; if that frame was U U, its pitch field held a hiss index, so the piece was played at a random pitch. Fixed by keeping the frame’s own pitch (UU_WO_GUARD, interp_wo()). [8]
Plate reverb (reverb is REE-verb): an old studio echo effect made with a big vibrating metal sheet. A listener’s description of a 20 ms hiss fade-out, which v3.8 then dropped. [8]
Plosive (say PLOH-siv): see stop consonant.
pYIN (say pee-YIN): a well-known offline pitch tracker, used as a reference to count Codec 2’s octave errors. Not perfect either. [2]
PN9: a fixed pseudo-random bit pattern that sotto mixes into the payload (scrambling). [6]
Postfilter (Codec 2): Codec 2’s --postfilter: a zero-bit trick that gives random phases to harmonics quieter than the background noise, so noise sounds like noise. [5]
Puncturing: deleting some of a code’s output bits in a fixed pattern, to make exactly the right number of bits. sotto’s code makes 3 bits per input bit and punctures down to 18 for every 7 (“rate 7/18”). [6]
Predictive quantizer: one that guesses each frame from the previous one and sends only the difference. Tried twice (the SottoV1 and SottoV2P classes in the code, the second with regular resets) and rejected both times, because a lost frame throws the decoder’s guesses off for several frames. [4]
Quantization (say kwon-tuh-ZAY-shun): rounding a number to one of a small set of allowed values, so it can be sent as a few bits. [4]
Quantization error: the difference between the true value and the rounded one. [4]
Quantization ladder: an experiment that starts with a perfect (all-oracle) vocoder and makes one piece realistic per step (made-up phases, then the LPC envelope, then rounding…). The drop at each rung shows what that piece costs. The code calls it the oracle ladder (sim/oracle_ladder.py). [7]
Quantizer (say KWON-tize-er): the rule for rounding. A uniform quantizer has equal steps (sotto’s energy); a log-uniform one has equal ratios between steps (sotto’s pitch). [4]
Rate-K envelope: the envelope sampled at K fixed frequencies (K = 20 in sotto), instead of at the harmonics. “Rate-K” as opposed to “rate-L”, the L harmonic amplitudes. to_rate_k(), from_rate_k(), KFREQ. [3]
Rayleigh fading (say RAY-lee): the rapid dips in signal strength in a moving vehicle, caused by reflections adding and canceling. [6]
RRC filter: “root-raised-cosine”. The filter that smooths the C4FM hops at the transmitter and matches them at the receiver. [6]
Regression (say rih-GRESH-un): two meanings. (1) In software, a bug that breaks something that used to work, like the v3.8 cleanup that lost the 50 Hz noise spacing. A regression test checks it can’t happen again. (2) In statistics, fitting a formula that predicts one set of numbers from another. A ridge regression adds a penalty that keeps the formula’s weights small so it doesn’t over-fit; v3.8’s vowel layer uses one. [8]
Resonance (say REZ-uh-nunce): a frequency where a space rings louder, like a bottle when you blow across it. Formants are resonances of the mouth; vowels also have one between about 4,300 and 7,300 Hz that moves with the vowel. [3, 8]
Sample / sample rate: one measurement of the microphone signal. sotto uses 8,000 per second; v3.8 takes in and plays out 16,000, for the hiss band. [1, 8]
Scrambling: mixing the bits with a known pseudo-random pattern (PN9) to avoid long runs of one symbol. Undone at the receiver. [6]
SDR: software-defined radio. A radio whose signal processing is done in a computer. The softmodem can use one. [6]
Shape: the 20-point envelope with its average removed, so only the ups and downs are left. [3]
Shape quantizer: the part of sotto that squeezes the envelope shape into 34 bits. Since v3 it’s a multistage vector quantizer (ShapeMSVQ) with 45% less error than before: 3 stages in sim/data/shape_msvq_34.npz up to v3.5, 4 smaller stages in sim/data/shape_msvq_9988.npz from v3.6. Up to v2 it was ShapeQ, a scalar quantizer: it took the DCT of the shape, gave each cosine-wave size its own uniform quantizer with a trained step size, and split the bits by greedy allocation (5, 4, 3, 3, 2, … bits). [4]
Shape measured / over the air: two rows on the v3.8 listening page. “Shape measured” uses a speaker shape taken straight from the original recording (the best possible); “over the air” uses the shape as the side channel delivers it. If the side channel works, they sound the same. [8]
Sibilant (say SIB-uh-lunt): a hissing fricative: “s”, “sh”, “z”, “zh”. Most of an “s” is above 4,000 Hz. The v3.8 sibilant rule sends a frame whose 4,000 to 8,000 Hz band is loud compared with its 0 to 1,000 Hz band as U U, so it gets a hiss index (sib_db; see z fix). [8]
Side channel: a slow, separate path for data inside the stream. In v3.8, one bit of every U U frame’s pitch field carries a 16-bit message per 2-second window (three 4-bit speaker numbers plus a CRC-4). [8]
Silence (true silence): since v3, energy code 0 tells the decoder to output nothing at all, instead of a faint buzz or hiss. Listeners had heard a “floor” of noise between words. [4]
sotto (say SOT-toh): the name of this vocoder. It’s an Italian word meaning “under”; speaking sotto voce means speaking quietly. [1]
sotto tracker: sotto’s joint pitch and voicing tracker (track_pitch_voicing() in sim/sinusoid.py). For every 10 ms frame it lists “unvoiced” plus several pitch candidates (Codec 2’s estimates nearby, doubled and halved), adds up six kinds of cost (bad fit, low pitch, disagreeing with Codec 2’s voicing, switching, jumping, missing odd harmonics), and uses the Viterbi algorithm to pick the cheapest path. Pitch is kept between 70 Hz (for doubled and halved candidates) and 500 Hz. v1 had a simpler pitch-only version, track_pitch(). [2]
sotto v0 to v3.8: versions of the sotto encoder, all with the same 48-bit, 20 ms frame. v0: the 20-point envelope. v1: the octave-correcting pitch tracker (SottoV0(pitch="tracked") in the code). v2: joint pitch and voicing tracking, and a 500 Hz pitch limit. v3: the trained vector quantizer and true silence. v3.1: the tail rule. v3.2: mixed voicing. v3.3: odd/even check. v3.4: concealment in the decoder. v3.5: lighter postfilter (1.35) and short concealment; SottoV35 in sim/vocoder_lab.py. v3.6: the smaller 29 KB codebook for the radio’s chip, and energy measured only over the envelope’s 200 to 3,700 Hz; SottoV36. v3.7: one more pitch candidate in the tracker, a third of Codec 2’s estimate (SottoV37). v3.8: the hiss band, with 16 kHz in and out (SottoV38 in sim/hiss_band.py; the C port’s default). [1, 2, 4, 6, 8]
Sine wave: the smoothest possible repeating wiggle, with a frequency, amplitude and phase. [2]
Sinusoidal model (say sye-nuh-SOY-dul): describing speech as a sum of sine waves (the harmonics). Codec 2 and sotto are both sinusoidal vocoders. [2, 5]
Softmodem: sotto’s software modem (modem/), which sends and receives System Fusion frames through a sound card or SDR. [6]
Soft bits: see LLR.
Speech Commands: a large collection of short spoken words from thousands of people (CC BY 4.0). Part of the training speech for sotto’s codebook, and the source of the counting clips used for testing. [4, 7]
Speaker shape: three numbers (from PCA) that say how one speaker’s and microphone’s high band differs from the average. Sent over the side channel; until the first message arrives, the decoder uses the average speaker. [8]
Springs: listeners’ word for the twangy sound when a decoder holds a buzzy vowel steady through a lost signal. v3.5’s short concealment avoids it. [6]
Spectrogram (say SPEK-truh-gram): a picture of sound with time across, frequency up, and loudness as darkness. [2]
Spectrum (say SPEK-trum; plural spectra): loudness at each frequency at one moment. [2]
Spreading (interleaving): shuffling bit positions so each vocoder frame’s bits are sprinkled across the whole 100 ms radio frame. Then one short fade damages a little of every frame instead of all of one frame. [6]
Sweep: an experiment that tries a range of values for one setting (like the postfilter strength) on many voices and keeps the best. [7]
Step size: the distance between two neighboring levels of a uniform quantizer. [4]
Stop consonant (plosive) (plosive is PLOH-siv): a consonant made by closing the mouth completely and then releasing the air in a burst: “t”, “p”, “k”, “d”, “b”, “g”. [8]
Sync: a fixed bit pattern at the start of each radio frame that the receiver searches for. [6]
Synthesis (say SIN-thuh-sis): the decoder’s last job: turning the model back into sound by adding up sine waves. [5]
System Fusion (YSF): Yaesu’s digital radio standard. sotto uses its C4FM signal and frame layout, carrying sotto voice instead of AMBE. [6]
Tail rule: a voicing clean-up from v3.1. A voiced frame with a weak harmonic fit that is more than 20 dB quieter than the last 300 ms is played as unvoiced, so the fading end of a “v” doesn’t buzz on as “uh”. unvoice_tails(). [2]
Tail-biting: a way of running a convolutional code on a block so the end wraps around to the start, wasting no bits on start-up and shut-down. [6]
Time index: the exact time in a clip, in seconds, where a listener hears a problem. Get it and repeat it back before analyzing anything: twice the team analyzed the wrong word without one. [8]
Training: tuning a quantizer’s settings by running it on lots of real speech. The old scalar quantizer trained step sizes and bit allocation (training_set()); the vector quantizer’s codebooks are trained with k-means on about 480,000 frames from more than 600 speakers. [4]
U U frame (all-unvoiced frame): a frame whose voicing code says both 10 ms halves are hiss. About 24% of frames. Since v3.8 their noise harmonics are always 50 Hz apart, so their 7 pitch bits carry a 6-bit hiss index and one side-channel bit instead. [2, 8]
Unvoiced: a sound made by hissing air with no vocal-cord buzz, like “s”, “f”, “sh”, or whispering. The hissing consonants are called fricatives. [2]
Viterbi algorithm (say vih-TER-bee): a method for finding the best path through a long chain of choices without trying every path. The receiver uses it to decode the convolutional code, and sotto’s tracker uses it to choose pitch and voicing. A fixed-lag Viterbi decides each frame after looking only a fixed time ahead (80 ms for sotto) instead of waiting for the end. [2, 6]
Vector quantizer (VQ): rounding a whole list of numbers at once, by picking the closest pattern from a codebook and sending its number. It can learn which combinations really happen in speech, so it wastes fewer bits than rounding each number separately. [4]
Vocoder (say VOH-koh-der): “voice coder”. A codec that sends a description of how a voice is being made, instead of the sound wave itself. [1]
Voiced: a sound with vocal-cord buzz, like vowels, “m”, “n”. [2]
Voiced fricative (say FRIK-uh-tiv): a fricative with vocal-cord buzz added, so it is buzz and hiss at once, like “v”, “z” or the “s” in “measure”. sotto sends these as mixed frames. [2]
Voicing: whether a frame is buzz, hiss or both. sotto sends 2 bits per 20 ms frame as one of four codes: U U, V V, M M, or an edge between U and M. Up to v3.1 the 2 bits were one yes/no for each 10 ms half. In the model it’s the voiced value: 0, 1, or 2 and up for mixed (the harmonic where the noise starts). [2, 4]
Voicing smoothing (vsmooth): a v3.8 encoder rule that turns short voiced/mixed/edge flickers inside a voiced run into mixed frames, which removed a stutter in the “zh” of “division”. [8]
Vowel layer: the quiet hiss v3.8 adds above 4,000 Hz under vowels, predicted in sixteen 250 Hz bands from the shape of the decoded 0 to 4,000 Hz spectrum, then corrected by the speaker shape (vowel_shape="fine"). [8]
Wo (“omega-zero”, said oh-MAY-guh ZEER-oh): the pitch written as an angle per sample: Wo = 2π × pitch ÷ 8000. sotto sends it with 7 bits, from 50 to 500 Hz in steps of 1.83% (400 Hz and 1.65% in v0 and v1). [2, 4]
Whisper rule: a v3.8 encoder rule. A run of at least 10 mixed frames with weak harmonics and a jumpy “pitch” is really a whisper, so it’s sent as U U instead of playing wandering harmonics. [8]
z fix: the v3.8 rule that sends any frame whose 4,000 to 8,000 Hz band comes within 10 dB of its 0 to 1,000 Hz band as U U with a hiss index, even if it’s voiced (sib_voiced=True, sib_db=-10). It fixed “prism” sounding like “prith-um”. [8]