Orbit Phone v1
Speech recognition trained from scratch in one night. 2.60 % word error rate on LibriSpeech test-clean.
Orbit Phone is the speech model of the Orbit family by Starlex AI. It is a 223M parameter Conformer with a CTC head, trained from random weights with no pretrained components, English first plus German and French. Apache-2.0.
| Parameters | 223M |
| Training | about 12 hours on 3 GPUs, about 230,000 hours of audio processed |
| Languages | English, German, French |
| License | Apache-2.0 |
Results
Word error rate, lower is better. Greedy CTC decoding, no language model, text normalized (lowercase, no punctuation).
| Test set | Orbit Phone | Whisper small (OpenAI, 244M) | Whisper base (OpenAI, 74M) |
|---|---|---|---|
| LibriSpeech test-clean (2,611 utterances) | 2.60 % | 3.4 % | 5.0 % |
| LibriSpeech test-other (2,932 utterances) | 6.35 % | 7.6 % | 12.4 % |
| MLS English test (3,769 utterances) | 9.45 % | ||
| VoxPopuli English test (1,830 utterances) | 7.38 % | ||
| MLS German test (3,394 utterances) | 7.62 % | ||
| VoxPopuli German test (1,957 utterances) | 10.27 % |
The Whisper numbers are OpenAI's published results. They are for orientation, not a head to head comparison, because Orbit Phone saw LibriSpeech training data and Whisper is evaluated zero-shot. Clips longer than 30 s were skipped in our evaluation.
Try it
pip install torch safetensors tokenizers soundfile numpy
python orbit_phone.py speech.wav
from orbit_phone import load_orbit_phone, transcribe
model, tok, feat = load_orbit_phone(".")
print(transcribe("speech.wav", model, tok, feat))
Everything needed to run the model is in this repository. No other library or checkpoint is downloaded.
How it was built
- Data. LibriSpeech (960 h), subsets of MLS English and German, VoxPopuli English and German, and subsets of YODAS-Granary English, German and French. Mix of about 60 % English, 30 % German, 10 % French.
- Tokenizer. Own BPE vocabulary with 1,024 pieces.
- Training. Muon (Newton-Schulz orthogonalized momentum) on the 2D block matrices and AdamW on the rest, with warmup, a constant phase and linear decay over the last 30 % of a fixed time budget. SpecAugment, bf16, 243,504 steps.
- Fixed batch shapes. Audio is sorted into 23 length buckets with about one hour of audio per step. Fixed shapes made training about 50 times faster than variable shapes on our hardware, from 353x to 19,000x real time.
Architecture
- Input: 16 kHz audio, 128 log-mel features at 100 Hz, normalized per utterance
- Encoder: 3 strided convolutions (8x down to 12.5 Hz), 16 Conformer blocks (d 768, 12 heads, RoPE self-attention, depthwise convolution kernel 31, macaron feed-forward)
- Output: CTC over the 1,024 piece vocabulary, intermediate CTC after block 8 during training
- Output text is lowercase without punctuation in v1
Limitations
- Proof of concept after one night of training. Not for safety-critical or legal transcription.
- No casing, punctuation, timestamps or language identification in v1.
- Weaker on rare names and German word endings. No language model in decoding.
- Audio longer than 30 s is cut into fixed 30 s windows, words at the cut can be lost.
The Orbit family
| Model | What it does |
|---|---|
| Orbit-Nano-Compliance-E2B | German and English compliance chat, rank 1 of 7 on our compliance benchmark |
| Orbit-Nano-Compliance-E2B-GGUF | The same model for laptops |
| Orbit Phone | Speech recognition |
| Orbit-Vision | Questions about images, charts and documents (research preview) |
| Orbit-Image | Pixel art from text (research preview) |
Data attribution and license
Weights and code Apache-2.0. Training data licenses are LibriSpeech CC-BY-4.0, Multilingual LibriSpeech CC-BY-4.0, VoxPopuli CC0 and YODAS-Granary CC-BY-3.0. We thank the authors of these datasets.
Built by Starlex AI, a Foundation AI Lab.
- Downloads last month
- 6
