Swanada shares speech-recognition models across 12 languages
Swanada: Split-Encoder Capability-Routed Conformer CTC and Sub-30 MB On-Device Speech Recognition Across Twelve Scripts · arXiv:2611.13814
In plain words
Swanada is a speech-to-text system that runs in a browser or local program. It recognizes Classical Sanskrit, Indian English, and ten other Indian languages. One shared part processes the sound, while a small language-specific part produces text in the selected script.
On 480 held-out studio utterances, the larger Studio version made fewer character errors than the smaller Mobile version. Public test results varied across the different speech collections.
The research
Running speech recognition locally can keep microphone audio off remote servers, but supporting many languages can exceed browser storage limits.
Swanada’s results make the trade-off more concrete: sharing a sound-processing model can shrink a twelve-language Studio fleet, while recognition accuracy still depends on which speech collection and model size are tested.
Swanada shares a speech encoder across twelve languages and switches native-script output heads, reducing the storage needed for a multilingual on-device system.
The system shares a sound-processing encoder across languages and selects a separate native-script output head for each one.
The evaluation measured character error rate, the share of reference characters that differ from recognized text: Studio averaged 5.43% on 480 held-out utterances, 8.87% on 960 public clips from eleven language tests and wild Sanskrit, and 12.06% across 1,280 clips in 16 public test sets.
The results show recognition on both held-out studio speech and varied public tests, while the changing aggregate reflects a broader mix of speech. The split design also reduced reported twelve-language Studio fleet storage compared with twelve full checkpoints.
The recognition figures describe the paper’s specified held-out and public test sets. The Mobile comparison changes both positional encoding and head calibration, so it cannot show which change accounts for the small CER difference.
What the evidence shows
- The deployment trade-off
Swanada is an open-source, client-side streaming speech-recognition system for Classical Sanskrit, Indian English, and ten Indian languages. The paper addresses a practical tension: separate full models take substantial storage, while one combined vocabulary can blur distinctions between writing systems.
Swanada’s design shares sound processing across languages while selecting language-specific output, and runs in browser WebAssembly, Node.js, and Python.
- Sharing cuts fleet storage
The decisive storage comparison is for a twelve-language Studio fleet: Swanada’s split encoder and language heads occupy 183.4 MiB, compared with 2,128.5 MiB for twelve monolithic Conformer-Large checkpoints. This shows a substantial storage reduction for that fleet.
It does not establish that all other model or runtime conditions were identical, or that the same storage figures apply to the Mobile tier.
- Recognition across speech sets
On the held-out studio suite, Swanada Studio’s mean character error rate was lower than on the broader open public tests, which combine different speech collections. Those aggregates cover distinct populations and do not imply equal recognition accuracy across their component sets.
- Mobile results have limits
The Mobile comparison on the held-out suite is much closer: CER was 18.56% for the configuration with dynamic positional encoding and calibrated heads, versus 18.58% for the static-position, uncalibrated-head configuration.
Because both settings change together, the comparison cannot attribute the difference to calibration alone. The authors also report difficult public-test cases for Mobile, including meeting speech and wild Sanskrit, which qualify broad claims about its robustness.
Deep
- Original paper authors
- Hemanth HM
- Submitted to arXiv
- License
- CC BY 4.0