Listen: real hardware examples¶

Eight sentences, two voices, each one played twice: once as it went in, once as it came back from a real AMBE codec chip.

Every "through the AMBE-3000" file on this page was produced by a DVSI AMBE-3000 vocoder chip — the silicon inside a NW Digital Radio ThumbDV — encoding the audio to 72-bit D-STAR frames and then decoding those frames back to sound. No software AMBE implementation was involved in producing the files on this page: the chip did the encoding and the decoding, and this site only measured it. No AMBE encoder or decoder appears anywhere in this repository.

How these were made¶

The chain is deliberately short and each stage is a file you can inspect:

1. Synthesis. Piper text-to-speech, one male voice (en_US-libritts_r-medium, speaker 690 of the 904 it carries) and one female voice (en_US-ljspeech-high), the same four sentences in both. Both voices come from corpora that can be redistributed under this site's licence, one public domain and one CC BY 4.0; the provenance table below names them, and docs/assets/audio/MANIFEST.md records the model checksums, the speaker choice and the exact synthesis parameters. Synthetic speech also means the material is identical between voices and free of room and microphone differences.
2. Format. Resampled to 8 kHz, 16-bit, mono — AMBE's native rate.
3. Encode. Pushed frame by frame into the ThumbDV, 160 samples in, 9 bytes out, per 20 ms. 4. Decode. Those same 9-byte frames pushed back into the same chip, 9 bytes in, 160 samples out.
4. Level. Each file scaled by one constant gain to −20 dBFS RMS (backed off if the peak would exceed −1 dBFS). No compression, no EQ, no noise reduction. The decoded files needed between −0.66 and +0.03 dB, so the codec had already preserved level well; the step exists so that an A/B comparison is about timbre and not about loudness.

Across the eight clips the hardware emitted 1056 frames, 9504 bytes, for 21.05 seconds of speech. At one 9-byte frame per 20 ms that is exactly 3600 bit/s, which is the full D-STAR voice frame: 2400 bit/s of vocoder data plus 1200 bit/s of error correction. — JARL D-STAR system specification; DVSI product documentation

The raw channel bytes are in the repository next to the audio, as docs/assets/data/frames/<clip>.ambe, together with the JSON that the animations on this site draw. Their exact contents and provenance are documented in docs/assets/data/SCHEMA.md.

What to listen for, in general¶

Three things are worth attending to on every pair, and each one is visible in the measurements as well as audible:

The pitch survives almost exactly. Tracking the fundamental through both files, the decoded pitch follows the original to within a median of 0.62–1.14 % across all eight clips, and a 90^th percentile of 1.4–4.3 %. A codec spending 3600 bit/s cannot afford to be wrong about pitch, and this one is not.

The envelope survives. Cross-correlating the loudness envelopes of the two files gives 0.967–0.991 at the matching offset on every clip. Syllables land in the same places with the same relative weight.

The texture does not survive intact. What changes is everything the harmonic model has to approximate: the exact shape of noise, the fine detail between harmonics, the attack of a consonant. The Multi-Band Excitation model represents a frame as harmonics of one fundamental, with each frequency band declared either voiced or unvoiced — Griffin & Lim 1988, and the places where that description is a poor fit are exactly the places where the pair sounds different.

Note what is not on that list: overall band level. Summed as energy over the energetic frames of a clip, the decoded audio comes back within 2.2 dB of the original in every 1 kHz band from 0 to 4 kHz, on all eight clips, and more often above it than below (19 of the 32 band figures are positive). The codec is not simply losing the top end. It is rebuilding it out of a different kind of signal, and the measurement that shows this is periodicity rather than level, which is the thread the clips below follow.

There is also a delay: the decoded audio arrives 31.6 to 34.4 ms after the original, which is between one and two frames. The players below are independent, so you will not notice it, but it matters if you ever line the two files up.

Male voice — en_US-libritts_r-medium¶

Median fundamental across these four clips: 132–153 Hz.

a) "The quick brown fox jumps over the lazy dog."¶

The stock pangram, here mostly as a baseline: 89 % of its energetic frames carry a detectable fundamental, so the harmonic model is on home ground and the two files track each other closely. Listen instead to the joins — the ck in "quick", the x in "fox", the j in "jumps". Those are the moments where the frame has to change its voiced/unvoiced description abruptly, and where the reconstruction has the least to work with.

b) "CQ CQ CQ this is a test of the AMBE voice codec."¶

The on-air case: a CQ call is the first thing anyone hears through this codec. 3.17 seconds of speech became 159 frames, 1431 bytes in total. This is the clip the animations elsewhere on the site are built from, so if a waveform or spectrum figure looks familiar, this is what it is showing. Listen to the hard K of each "CQ": the burst is present but blunted, because a click is the least harmonic thing a voice does.

c) "She sells sea shells by the sea shore."¶

This one is a stress test, and it is the hardest clip on the page for the model: only 66 % of its energetic frames carry a detectable fundamental, the lowest figure of the eight. sh and s are pure noise, and the model's answer is to declare those bands unvoiced and fill them with noise rather than harmonics.

The interesting part is what that does and does not cost. Over the 31 energetic frames of this clip with no detectable fundamental, the decoded audio comes back within 1.5 dB of the original in every 1 kHz band — the level survives. What does not is the structure inside it: a band-by-band periodicity measurement of the decoded audio on those same frames averages 0.10 or below in all eight bands from 0 to 4 kHz on a 0–1 scale, against 0.63 in the lowest band on this clip's pitched frames. The sibilants are not being attenuated; they are being re-manufactured from a noise source at roughly the right level. Listen for the difference between "sells" and "shells": the distinction is reconstructed rather than reproduced.

d) "We were away a year ago."¶

The opposite test: a sentence with no stops and no fricatives, where every one of the 54 energetic frames — 100 % of them — carries a detectable fundamental. This is the harmonic model at its best — a single fundamental with a smoothly moving set of harmonics is exactly what the codec is built to carry — and it is the pair on this page where original and decoded sit closest together.

Female voice — en_US-ljspeech-high¶

Median fundamental across these four clips: 174–219 Hz, about 45 Hz above the male voice. That matters to a harmonic codec: at 135 Hz there are twenty-nine harmonics below 4 kHz to describe, at 180 Hz there are twenty-two. The same bits are spread across fewer, more widely spaced components. — Griffin & Lim 1988

a) "The quick brown fox jumps over the lazy dog."¶

The same sentence as the male pangram above, which makes the two directly comparable: play lr-a and this one back to back and the codec's character — what it keeps, what it smooths — shows up as the thing they have in common, rather than as a property of one voice.

b) "CQ CQ CQ this is a test of the AMBE voice codec."¶

4.16 seconds, 208 frames, 1872 bytes. Summed over this clip's energetic frames, the decoded spectrum sits within 0.2 dB of the original in the 0–1 and 2–3 kHz bands, 2.2 dB below it in the 1–2 kHz band and 1.9 dB above it in the 3–4 kHz band. The round trip is close to level-neutral across the whole band; what it does not preserve is what that energy is made of. Listen to the word "codec" itself, where a hard k is followed immediately by a voiced vowel.

c) "She sells sea shells by the sea shore."¶

The clearest demonstration on the page of the voiced/unvoiced split. On this clip's sibilant frames — the energetic ones with no detectable fundamental — a band-by-band periodicity measurement of the decoded audio averages 0.13 or lower in all eight bands from 0 to 4 kHz on a 0–1 scale. The output there is noise, with next to no periodic structure. The same measurement averages 0.85 in the lowest band across this clip's pitched frames. That is the same story as the male "shells" clip, in a voice whose sibilants are stronger to begin with.

d) "We were away a year ago."¶

94 % of this clip's energetic frames carry a detectable fundamental — only four do not. At 1.58 seconds it cost 720 bytes. Take the sentence as the answer to "what does 3600 bit/s actually buy you", because on material this cooperative the answer is: very nearly everything except the timbre.

Provenance summary¶

Both voices were chosen for their licence as much as for their sound. A Piper voice inherits the terms of the corpus it was trained on, and several of the popular English voices are trained on datasets carrying a non-commercial or share-alike condition that cannot be redistributed under this site's CC BY 4.0 grant. Neither of these two does. The female clips come from a public-domain corpus and carry no upstream condition. The male clips come from LibriTTS-R, which is CC BY 4.0 and requires a credit: speech synthesised with a Piper voice trained on LibriTTS-R (OpenSLR 141), used under CC BY 4.0. That credit is repeated in Sources and method, and anyone who reuses an lr-*.wav file takes the requirement with it. The fine-tuning lineage named in the table above is a separate and unsettled question; docs/assets/audio/MANIFEST.md states it and makes no claim about how far it reaches into a synthesised waveform. tools/make-audio.sh records the licence rule at the top of the file, so that a future substitution does not quietly reintroduce a voice that cannot ship here.

Nothing on this page was produced by, or compared against, a software AMBE implementation. This project is independent of Digital Voice Systems, Inc.; AMBE is a DVSI trademark, used here only to name the technology under discussion.

Next: The patent landscape. Which patents cover what, which two are still in force, and what that constrains for someone building an open implementation. Previously: What is not in the public record.

Last reviewed: 2026-08-11.