Software Automatic Mouth: the 1982 Commodore 64 formant speech synthesiser, as an instrument.
Type some text, then either play it from MIDI (each note speaks the next word, sung at the
pitch of the key) or let the host transport read it aloud (one word per step from a start bar,
locked to the bars and beats). See docs/design.md.
SAM's code is vendored in third_party/sam and has no licence (abandonware; see
third_party/sam/SAM-LICENSE.md). It carries local fixes, listed in third_party/sam/PATCHES.md.
No audio input. Stereo output (the voice is mono, sent to both channels). One MIDI input.
| Mode | What plays | Pitch | Needs |
|---|---|---|---|
| Keys | A note speaks one word; Advance picks which: Next word cycles through the text (and starts over when the transport starts), Key = word speaks word (note - 48) modulo the word count |
The key, tuned (see below) | MIDI |
| Transport | Word n starts at step n counted from Start bar; Loop repeats the text. A word entered mid-step (after a locate or a loop wrap) is spoken from the matching point | Pitch, or the fixed Monotone pitch | Host transport. Silent while stopped; Speak plays one pass without it |
Words are the whitespace-separated pieces of the text. A lone _ is a rest and still takes its slot.
Read it as chooses English (SAM's own spelling rules; . and ? shape the pitch) or
Phonemes (/HEH3LOW2 WER4LD: SAM phoneme codes, digits 1-8 for stress, /H a breath).
| Control | Range | Default |
|---|---|---|
| Mode | Keys / Transport | Keys |
| Speed | 20-200 (lower is faster) | 72 |
| Pitch | 16-255 (SAM units; Transport mode) | 64 |
| Mouth, Throat | 0-255 | 128, 128 |
| Monotone | Off / On (fixed pitch, as SAM's -sing; Transport mode) |
Off |
| Phonetic | Off / On | Off |
| MIDI In | Off, Ch 1-16, All | All |
| Advance | Next word / Key = word | Next word |
| Transpose | -24 to +24 st | 0 |
| Release | 5-1000 ms | 60 ms |
| Step | 1/16, 1/8, 1/4, 1/2, 1/1, 1 bar, 2 bars, 4 bars | 1/4 |
| Start Bar | 1-256 | 1 |
| Loop | Off / On | On |
| Speak | trigger | |
| Tone | 1-12 kHz low-pass, top is open | 8000 Hz |
| Level | 0-1 | 0.70 |
| Current Word | read-only status: slot being spoken, 0 when idle | |
| MCP Port | read-only status: port of this instance's MCP server, 0 when off |
The panel also has the six voices from the SAM manual (SAM, Elf, Robot, Stuffy, Old Lady, Alien) as presets of Speed, Pitch, Throat and Mouth.
MIDI: note on/off (velocity scales level), pitch bend (+-2 semitones), CC 120 and 123 release everything. Eight voices, the oldest is stolen.
The text is saved with the project as the text state (up to 2000 characters, 512 words); every other
setting is a parameter.
An embedded MCP server (loopback, port 7230 and up) lets an agent or script set the words or phonemes,
speak them now, and move the parameters while the host plays. From a shell:
scripts/sam-say.sh hello world. Disable with DOWNSPOUT_SAM_MCP=0. See docs/mcp.md.
- Tuning. SAM sets pitch as a whole number of ticks of 3.24 samples at 22.05 kHz, so the nearest pitch can be well off a key: up to about 0.7 semitone at C5 and about 2 semitones near B6. Each note is rendered at the nearest pitch and resampled the rest of the way; measured over C2 to B6 the result is within 12 cents of the key. The resampling moves the voice's formants by the same ratio, so the top octave takes on a slightly smaller, brighter voice.
- Sound. SAM is 8-bit at 22.05 kHz. Tone softens the aliasing; set it to open for the raw sound.
- CPU. A word is rendered by SAM on the audio thread when it starts: about 30-50 us for a short word, roughly 1 ms per second of speech. A very long word (tens of characters) takes proportionally more. Nothing is allocated.
- Limits. One word is at most 48 characters. A word that runs past its step is cut by the next one, with a few milliseconds of fade.
Part of the root build (DOWNSPOUT_BUILD_SAM, default ON):
cmake --build build --target sam-vst3
ctest --test-dir build -R sam
Run env -u DISPLAY scripts/capture-plugin-screenshots.sh sam for the catalogue image.