Streaming Speech-To-Text and Text-To-Speech that run natively in the Surogate engine, on a CPU or a GPU.
Surogate Speech is our line of small, open speech models for agents: a model family per job, Jackrabbit to listen and Amami to speak, shipped language by language.
Weights under CC-BY-NC-4.0 for research and non-commercial use; commercial licences through Invergent.
A 110M-parameter voice that runs in real time on two CPU cores, no GPU. It puts the stress where native speakers put it, says the English words Romanians use at work (meeting, call, update, backup) the English way, reads numbers, dates and amounts in words, and dictates codes, IBANs and phone numbers one character at a time from recorded clips, so a voice agent never reads a code wrong.
Seven lines, two voices, one bank support call: English words, an SMS code dictated from recorded clips, an amount in lei and bani. Generated on a laptop CPU with 2 threads, first audio in about 40 ms per line.
word error rate on a set of ordinary Romanian sentences, transcribed back by the Surogate Romanian speech recogniser
Model cardA 116M-parameter recogniser that writes cased, punctuated text, from a file or live over HTTP or WebSocket. Below, real people reading real sentences.
Press play. Words appear as the recording reaches them.
Partials: Jackrabbit Streaming run over this recording, shown at the moment of audio it had heard. Final: from the published evaluation run, re-read with full context and a 4-gram LM after the pause; words it changed are underlined.
What was saidPorțiuni mari mai spre nord sunt destul de puțin populate, iar unele sunt aproape sălbăticie nepopulată.
Porțiuni mari, mai spre nord, sunt destul de puțin populate, iar unele sunt aproape sălbăticie nepopulată.
Porțiuni mari, mai spre nou, sunt destul de puțin populate, iar unele sunt aproape sălbaticie nepopulată.
Porțiuni mari, mai spre nori, sunt destul de puțin populate, iar unele se aproape să îl pătricie nepopulată.
Clips picked to show the difference. Over all 883 FLEURS Romanian test clips the word error rate is 5.69% for Jackrabbit (CTC + 4-gram), 5.95% for the 1B model and 8.42% for the 1.55B model, each run through the same harness; every transcript is in the evaluation dataset. Recordings: FLEURS by Google, CC BY 4.0.
Speech recognition · Jackrabbit 110MWrites cased, punctuated text from recordings. A 116M-parameter model that runs comfortably on a CPU, with no Python in the serving path.
Live speech recognition · Jackrabbit 110M StreamingShows words as you speak, then writes a final, punctuated sentence when you pause. Live audio over HTTP or WebSocket.
From the model cards. Amami runs on CPU only and needs Surogate 1.5.5 or later; drop --gpus all to run Jackrabbit on a CPU too.
hf download surogate/amami-110m-ro --local-dir amami-110m-ro
SUROGATE_TTS_PYTHON=/path/to/venv/bin/python surogate serve --tts amami-110m-ro --port 8080
curl localhost:8080/v1/audio/speech -H "Content-Type: application/json" \
-d '{"input": "Bună ziua! Codul de confirmare este 4B7X9.", "voice": "male"}' -o out.wavdocker run --gpus all -p 8000:8000 ghcr.io/invergent-ai/surogate:1.5.4 \ serve --stt surogate/jackrabbit-110m-ro-streaming --host 0.0.0.0 --port 8000 curl http://localhost:8000/v1/audio/transcriptions -F [email protected]
Jackrabbit listens and Amami speaks, from one container, on a CPU or a GPU.