Launa: Sovereign AI Voice Models from Salesforce
Introducing Launa Omni v1 and Launa Transcribe v2
LAUNA / VOICE AI
MODEL CARD
Launa in comparison: in comparison with models
| Benchmark | Launa Omni v1Salesforce | Grok Voice 2.0xAI | Inkling (large)Thinking Machines | GPT-Live-1OpenAI | GPT-Realtime-2OpenAI |
|---|---|---|---|---|---|
| General voice benchmarks | |||||
| VoiceBenchSpoken instruction & safety | Best reported score: 90.60% | 85.8% | 76.6% | 79.2% | 78.2% |
| Audio MultiChallengeNatural multi-turn dialogue | Best reported score: 59.07% | 35.2% | 46.2% | 28.3% | 30.5% |
| Service agents & tool calling | |||||
| Tau VoiceCustomer-service task completion · Pass1 | Best reported score: 62.3% | 60.7% | 58.80% | 52.9% | 34.4% |
| EVA BenchCustomer-service accuracy & experience · mean score | 0.70000.6449 / 0.7551 | Best reported score: 0.73850.7203 / 0.7566 | 0.62960.6879 / 0.5715 | — | 0.67260.6010 / 0.7456 |
| Multilingual tool callingSpoken function calling · Speech-BFCL | Best reported score: 72.9% | 64.8% | 65.8% | 55.1% | 65.2% |
| Employee agents · internal Slack | |||||
| Slackbot VoiceWorkplace assistant tasks · internal eval | Best reported score: 96.5% | 50.0% | 49.2% | 82.7% | 85.4% |
Best reported score per row
Launa Omni v1
— Not reported
Results reported by the Launa team. EVA covers clean customer service only; the smaller values show EVA-A / EVA-X. Scores use each benchmark’s own scale.
| Benchmark | Launa Transcribe v2Salesforce | Ink 2Cartesia | Voxtral SmallMistral AI | Universal-3.5 ProAssemblyAI | Nova-3Deepgram | Granite Speech 4.1 2BIBM | Voxtral MiniMistral AI | Parakeet CTC 1.1BNVIDIA | Canary 1B v2NVIDIA | Whisper large-v3OpenAI |
|---|---|---|---|---|---|---|---|---|---|---|
| MU-Bench EnglishMeaning-changing UER · 817 English utterances | 2.57% | 8.20% | 2.33% | Best score: 2.08% | 3.06% | 5.63% | 3.55% | 5.75% | 5.39% | 6.85% |
| HALAS corrected testCorrected test references | Best score: 28.70% | 30.73% | 30.03% | 29.43% | 36.42% | 59.01% | 32.74% | 40.16% | 32.71% | 60.40% |
| FLEURS EnglishEnglish test split | 3.33% | 4.72% | 2.67% | Best score: 2.55% | 4.97% | 3.93% | 3.37% | 5.30% | 4.47% | 4.01% |
| VoxPopuli-Cleaned-AA628 English clips | Best score: 1.66% | 2.06% | 2.28% | 2.22% | 2.95% | 4.53% | 2.44% | 3.43% | 4.62% | 3.00% |
| Earnings22-Cleaned-AA6 earnings calls | 4.94% | Best score: 4.77% | 5.94% | 7.11% | 10.94% | 9.96% | 6.55% | 12.19% | 8.28% | 8.09% |
| Monsoon Open ASR en-INIndian English public test | Best score: 4.17% | 5.03% | 4.43% | 4.21% | 5.62% | 5.49% | 5.96% | 6.17% | 6.03% | 4.80% |
| AMI IHM testIndividual headset microphone | 14.00% | 11.54% | 18.06% | 15.03% | 16.62% | Best score: 10.31% | 18.91% | 18.52% | 17.94% | 18.31% |
Best score per row
Launa Transcribe v2
↓ Lower WER / UER is better
Scores follow the public Artificial Analysis methodology: reference and hypothesis are normalized, WER is computed per clip, then averaged weighted by audio duration. Monsoon en-IN reports duration-weighted WER rather than the dataset’s corpus WER.
THE TEAM
Salesforce AI Research - Resonance
Launa was trained in partnership with Arcanum.
- Jason Lee
- Alexey Dyakov
- Ben Stolovitz
- Daniel Lee
- Jianing Yang
- Kah Mun Chia
- Patrick Maynard
- Raphael Shu
- Ritvik Gupta
- Shay Scerri
- Tanay Dixit
- Young Mo Kang
- John Emmons
- Silvio Savarese
- Samir Walji
- Ansh Moraje
- Daniel Rupawalla
- Michael Crystal
- Theo Chapman