Learn/Remade with RustDigital Freedom

The Mercury Speech Engine: Less Memory Traffic, No Verified Speedup Yet

The Mercury speech engine in FFAI now memsets 3.3 MB instead of 573.3 MB on one ASR test set. It is unreleased work on master, and its timing run was ruled inadmissible.

Signed by M·
A small wooden radio and an open lined notepad with a pencil across the page on a light wooden counter in soft morning light, a potted plant behind
Did the Mercury speech engine get faster in November 2026?

No speedup is verified yet. November work on the FFAI master branch cut counted memory traffic, for example ASR memset from 573.3 MB to 3.3 MB on 13 LibriSpeech clips, but the same-day timing run was recorded as inadmissible, and the work is not in the published ffai-mercury 1.1.0 crate.

The Mercury speech engine is FFAI's speech component: Whisper-class speech recognition and VITS-class text to speech, in Rust, on the CPU. This month it got a development update that is genuinely interesting and genuinely unfinished, and the right way to tell it is as a memory-traffic story, not a speed story.

On one test set, the Mercury speech engine stopped writing 570 megabytes of zeros it never needed. Speech recognition memset fell from 573.3 MB to 3.3 MB. Bytes copied into the attention cache fell from 510.0 MB to 14.4 MB. Those are real cuts. The same day, the team tried to time the change, and the ledger recorded that timing pass as inadmissible. So there is no faster-transcription claim on this page.

The work also is not released. Commits bae28c23 and ef6c355f landed on the FFAI master branch after the published ffai-mercury 1.1.0 crate. If you install Mercury from crates.io today, you get 1.1.0, which does not contain any of this.

Our earlier piece, Mercury, pure-Rust Whisper without Python, covers what Mercury is and its earlier comparisons. Those older numbers were not re-measured here and stay on that page. The product page is Mercury on MATA, and the builder notes are at remadewithrust.com.

What the November Work Removed

The campaign counted bytes and calls, not seconds. The inputs were 13 English clips, 12 from the LibriSpeech test-clean set plus one long-form clip, 176.3 seconds of audio and 669 generated tokens on the whisper-tiny.en model. Text to speech used 8 Harvard sentences with the piper lessac-medium voice, seed 7. LibriSpeech is a public read-speech corpus, which makes it a fair, repeatable test bed.

  • Speech recognition memset: 573.3 MB to 3.3 MB.
  • Attention-cache bytes copied: 510.0 MB to 14.4 MB.
  • Vocabulary heap: 142.5 MB to zero.
  • Allocations: 914,722 to 463,103.
  • FFT calls: 31,000 to 18,972, because all-silent audio frames now skip the transform.
  • TTS memset: 502.6 MB to 34.4 MB.
  • TTS tensor copies: 232.0 MB to 126.0 MB.
  • Beam-search elements sorted: 17.9 million to zero, because the top choices are now selected instead of sorted.

The counters ran on Windows with Rust 1.95 on an x86-64 machine with AVX2 and F16C. The CPU model is not named. Every number is repository-reported, and we reran none of it.

Why There Is No Speed Claim

Less memory traffic usually helps. Usually is not a measurement, and this campaign's own records say why the clock stayed silent.

The Mercury Timing Run Was Inadmissible

The team ran an alternating before-and-after timing pass on an Intel i7-14650HX, 12 runs per binary per arm. The result was noise. Two runs of the very same binary differed by more than the before-and-after gap, and speech-recognition wall time drifted from about 20 seconds to about 4 seconds across the session. The ledger labels the pass inadmissible and keeps changes on the counter plus an output-identity check instead. We follow the ledger. There is no delivered latency number for this work, faster or slower.

Bytes Moved Are Not Resident Memory

It is tempting to read 573.3 MB to 3.3 MB as "Mercury now needs half a gigabyte less RAM." It does not say that. Memset bytes are summed over the whole run. A program can write the same small buffer thousands of times and rack up a large total while holding very little at once. The ledger itself estimates the removed bytes at roughly 0.1 seconds at memory bandwidth, smaller than the one to two seconds of noise between identical runs.

Older Comparisons Stay Where They Are

The older comparisons in our Whisper article and the README's older text-to-speech figures were not part of this campaign. Nothing here replaces them, and nothing here adds a new comparison against other engines. The Whisper model family is described in the original Whisper paper; Mercury is a Rust implementation of that class of model, not Whisper.cpp.

What It Means for a Private Voice at Home

A microphone in the house hears more than commands. Where that audio is processed is the privacy question, and keeping it local only works if a home machine can afford it.

Identical Output as the Gate

Each change was kept only if counters fell and output stayed identical on the harness. Greedy and beam-search transcripts matched the baseline byte for byte. The same ledger also found that text to speech, run 150 times per binary, produced exactly two waveforms, before and after the campaign alike. The README still says the same text and seed give a byte-identical WAV. The ledger's narrower wording is the accurate one: the code path is deterministic, and not every process run was. Printing the awkward finding beside the good one is what the Digital Freedom principles ask of any claim.

A Voice That Does Not Have to Leave

The Federal Trade Commission's 2023 complaint about Alexa voice recordings describes recordings and transcripts kept long after they were needed. Speech recognition that runs on your own hardware sidesteps that retention question for the audio it handles. Rust also helps with a different risk: CISA's case for memory-safe roadmaps explains why memory-safe languages remove a whole class of bugs. The NIST Privacy Framework is the broader reference for managing processing risk.

What to Watch Before It Ships

For now, the Mercury speech engine work is a development update. Watch for a new ffai-mercury version and a changelog entry before treating any of it as installed. An environment switch, FFAI_VEINS=off, keeps four legacy paths for comparison. The long-term incentive is simple: a home companion that never phones home needs speech that is cheap to run locally. Local AI without uploading your life makes that case, the Freedom Guide gets you started, and Learn holds the rest.

FAQ

Short answers already in this article.

Is the November Mercury work inside ffai-mercury 1.1.0?

No. ffai-mercury 1.1.0 was published on 31 October 2026 from an earlier commit. The campaign commits bae28c23 and ef6c355f sit on FFAI master, the crate version still reads 1.1.0, and the changelog's Unreleased section is empty.

Does 573.3 MB to 3.3 MB mean Mercury uses less RAM?

No. Those figures are bytes written by memset, summed across a whole test run. They describe work the program stopped doing, not how much memory the process holds at once.

Why was the Mercury timing run thrown out?

Two runs of the same binary differed by more than the before-and-after gap, and ASR wall time drifted from about 20 seconds to 4 seconds during the session. The ledger estimates the removed bytes at about 0.1 seconds of memory bandwidth, below that noise.

Did the Mercury transcripts change after the copy cuts?

On the campaign's harness, greedy and beam-search outputs were byte-identical to the baseline. That identity holds for those runs and that clip set. It is not a word-error-rate study.

Is Mercury text to speech byte-identical for a given seed?

Not in every process. In 150 text-to-speech runs per binary the ledger found exactly two waveforms, the alternate appearing in 5.3 percent of runs before the campaign and 2.7 percent after, with no significant change. The ledger says the cause is open and predates this work.