Express series

Express is a series of three small speech and sound models built by Ducky Software, one of the four series under the Falcon name. Where the LIM, Waymark and Response series read text, sketches and tracks and return labels, answers and short replies, Express works in audio: a description of a sound into a short rendered cue, a line of text into speech, and a short recording into text. The two models are documented under the same nine-chapter contract as every Falcon model, carry the same status vocabulary, and are reached through the same Falcon API. Both are Released: Express Cue serves behind /v1/sound and Express Voice behind /v1/speak.

What Express is #

Three families cover three jobs. Sound models read a short description of a sound and drive a deterministic synthesis engine to render a cue of at most six seconds — a chime, an alert, a swoosh, a texture such as rain or static. Voice models read a short line of English text and return it spoken in one of a small fixed catalogue of voices. Transcription models read a short English recording and return the words as plain text. Nothing in the series clones a voice, plays a recording back, identifies a speaker or generates music.

Each model carries a family, a role among its siblings and a status, exactly as on the Falcon side; Families and Status & versioning define the vocabulary. Every Express model returns audio as a base64-encoded WAV inside the usual JSON envelope, with the duration and the identifier of the model that answered.

Sound #

Express Cue is the one Sound model. It reads a description of up to 400 characters and predicts a kind — one of eighteen, from tone and chime through swoosh and riser to rain, thunder and footsteps — plus a set of continuous synthesis parameters: duration, base frequency, partials, inharmonicity, an ADSR envelope, a noise colour mix, a swept low-pass filter, amplitude modulation, a pitch glide, repeats, reverb and gain. A fixed signal-processing engine renders those parameters to 24 kHz mono WAV of between 0.1 and 6 seconds. The parameters come back with the audio, so a host can edit and re-render them, and a seed makes a render repeatable byte for byte. The kind catalogue is listed on Output vocabularies.

Voice #

Express Voice is the one Voice model: seven consistent English voices, male and female, from one GPU service. Every voice is a fine-tune of SpeechT5, an open encoder–decoder speech model, with the HiFi-GAN vocoder published alongside the base producing the 16 kHz waveform. The voice lj, the default, is a 12,000-step fine-tune on LJSpeech, a public-domain corpus of about 24 hours of read English from one speaker; the six voices female-1 to female-3 and male-1 to male-3, added on 2026-09-14, are a second fine-tune of the same base, 15,000 steps, on six narrators of the Hi-Fi Multi-Speaker English TTS corpus, audiobook narration under CC BY 4.0. One speaker x-vector per voice is computed before training and stored with the weights; nothing is computed from audio at request time. It is Released and serves all seven entries of the catalogue that GET /v1/voices publishes. Its Evaluation chapter records the figures per voice and the earlier multi-voice sibling that was trained, evaluated and withdrawn.

Transcription #

Express2Text is the one Transcription model: a 20M-parameter conformer with a CTC head over 29 characters, trained from scratch — no pre-trained weights — on LibriSpeech, 960 hours of read English audiobook speech. It reads WAV, FLAC or OGG of at most 60 seconds at any sample rate, resamples to 16 kHz, and returns lower-case text without punctuation or numerals as digits. It is co-hosted on the Express Voice GPU service behind /v1/transcribe, so it shares that service's warm instance and answers in well under a second for a typical clip. The Express2Text page records its architecture, its training recipe and its word-error rates.

How the three relate #

The three models share the envelope, the audio quota bucket, the reference client's behaviour — a 45 s timeout, one retry after 2.5 s on a 503 — and the licensing stance below, and nothing else. Express Voice answers /v1/speak with text and an optional voice id from its catalogue of seven, lj by default, and returns 16 kHz mono WAV within a 600-character limit. Express Cue has its own route, /v1/sound, its own output rate of 24 kHz, and no voice at all. Express2Text goes the other way, audio in and text out, on /v1/transcribe, from the same GPU service as Express Voice.

A host that speaks and signals typically uses both: Express Cue for the sound that marks an event, Express Voice for the words that follow it. The home page shows where the series sits beside Falcon's four steps; API conventions documents the three audio routes with the rest of the Falcon API.

Licensing #

Each Express model records its base, its corpus and its licences on its page. In summary:

The six Hi-Fi voices of Express Voice are trained on the Hi-Fi Multi-Speaker English TTS corpus, audiobook narration released under CC BY 4.0; that attribution is carried here and on the Express Voice page, not in the API payload, and a host that redistributes audio in those voices keeps it. Express2Text is trained on LibriSpeech (Panayotov, Chen, Povey and Khudanpur, 2015), also CC BY 4.0, attributed here and on its page. No Express model is trained on recordings of users, and no Express model distributes its weights during the private preview.

Runtime and access #

Express Cue runs as a CPU service (0.2–0.7 s per cue when warm). Express Voice runs, since 2026-09-16, as one GPU service on an NVIDIA L4 that holds both fine-tunes and is kept warm with one minimum instance: a 14-word sentence renders in about 2.2–2.5 s and a short line in about 0.7 s, with the Falcon API passing the voice id through. Each runs behind routes of its own — a health route, its model route, and for Express Cue a direct render route — and is reached through the Falcon API, which meters the audio routes in the audio preview quota bucket. Both serve today; each page says in Runtime & deployment where its endpoint is. Access lists the routes a preview key covers.