Ofelia
Norwegian BokmålKyutai DSM · Variant C tokenizer · alignment-gated
● Voice research · Kapllan AIOslo · Stockholm
The voice, written down.
Public voice models are English-centric. Kapllan AI builds streaming STT, streaming TTS and voice-cloning models that start from Norwegian Bokmål and Swedish. Owned voices, on private compute.
§ Listen
Ofelia, in three voices. Press play.
Ofelia
Norwegian Bokmål · variantSame backbone, different speaker conditioning
Ofelia
EnglishCross-lingual baseline from the same checkpoint
Ofelia
Live capture · 2026-03-19Recorded from the production voice surface
Talk to Ofelia
Idle52 / 600 chars
Samples are raw outputs from Ofelia’s Kyutai DSM LoRA pipeline. The “Talk to Ofelia” box proxies your text through Kapllan’s private cluster — if the upstream voice service is offline the status line will say so instead of failing silently.
§ The voice stack
One production voice. Several research lines beneath it. Each one earns its place.
Ofelia
LiveProduction voice brain
The voice that powers the Kapllan voice surface. 48 speakers, ~184 ms first-token latency, runs on node3 behind the Unmute STT/TTS stack with a gpt-oss-20b brain.
Kyutai DSM (STT)
ResearchStreaming speech-to-text
Adaptation of kyutai/stt-1b-en_fr for Norwegian Bokmål (primary) and Swedish (secondary). DSM alignment-gate enforced as a preprocessing hard requirement — every (audio, text) pair must pass alignment_quality_gate.check_utterance().
Kyutai DSM (TTS)
ResearchStreaming text-to-speech
Adaptation of kyutai/tts-1.6b-en_fr for Nordic languages. Variant C tokenizer (8k unigram SentencePiece on EN+NO Wikipedia) achieves 0 % OOV at 1.94 tpw density.
VoxCPM 2.0
ResearchVoice cloning · 30 languages
2 B parameters, 48 kHz output. Used for design-mode and ultimate-mode cloning. LoRA + full-finetune pipelines wired through the kapllan-voice CLI; OpenAI-compatible /v1/audio/speech endpoint.
Pocket-TTS
ResearchOn-device TTS
kyutai/pocket-tts targeted at edge / on-device synthesis. Same alignment-gate constraints as the streaming pipeline.
§ How we work
Three constraints we choose, so the system stays honest.
Nordic-first
Norwegian Bokmål and Swedish, not English-translated.
Public voice models are English-centric. We build for the languages we actually speak. The corpus, the tokenizer and the alignment thresholds all start from Nordic baselines.
Alignment as a gate
Preprocessing is enforced, not learned.
Every (audio, text) pair we keep on disk has passed an alignment quality gate. Misaligned data does not enter training. This costs us throughput; it buys us stability.
OpenAI-compatible serving
Speech endpoints look like every other model on Kapllan AI.
The kapllan-voice FastAPI service exposes /v1/audio/speech so any client that already speaks OpenAI's audio API can route through Kapllan AI's private cluster without a rewrite.
§ What runs
Backbones, languages, where it serves from.
Backbones
Kyutai DSM (STT, TTS) · VoxCPM 2.0 · Pocket-TTS
Tokenizer
Variant C — 8k unigram SentencePiece, EN + NO Wikipedia corpus
Languages
Norwegian Bokmål (primary), Swedish (secondary), 30-language VoxCPM coverage
Serving
FastAPI · /v1/audio/speech · Unmute frontend · gpt-oss-20B brain
Production host
node3 · 10 TTS + 5 STT engines · Hetzner nginx → WG → 10.100.0.10
Research host
nxai-voice-lab + kapllan-voice-lab — Kyutai adaptation, VoxCPM cloning + finetuning
§ Adjacent
Voice is one of several capabilities.
Each Kapllan AI surface ties back to the same private compute and the same model ladder. Voice meets Llana in the chat; Ofelia speaks; K—1 reads.