Berget
Berget is a Swedish cloud provider. OSTT uses Berget for Klang Pianissimo, Swedish and Norwegian optimized Whisper models, and general-purpose Whisper Large V3, with processing hosted on European infrastructure.
Berget documentation:
- Models overview
- API reference
- Realtime transcription API
- Pianissimo realtime launch and file-transcription example
Models
| Model ID | Notes |
|---|---|
berget/klang/pianissimo | Swedish Parakeet-based model by Klang AI, published as KlangAI/pianissimo-sv on Hugging Face. Uses Berget's realtime API. Available since OSTT 0.0.27. |
berget/KBLab/kb-whisper-large | Swedish-optimized KB Whisper Large. KBLab reports 50,000+ hours of Swedish speech training and 47% average WER reduction versus OpenAI Whisper Large V3 across FLEURS, CommonVoice, and NST. |
berget/NbAiLab/nb-whisper-large | Norwegian-optimized NB-Whisper Large. NbAiLab reports 66,000 hours of training data and support for Norwegian, Bokmal, Nynorsk, and English. |
berget/openai/whisper-large-v3 | General-purpose multilingual Whisper Large V3. |
Check Berget's pricing for current model-specific rates.
Pianissimo params
Pianissimo requires OSTT 0.0.27 or later; older releases will report an unknown model. Authenticate with ostt auth login, choose Berget, then use ostt model to select Pianissimo.
ostt transcribe recording.mp3 -m berget/klang/pianissimo
ostt record -m berget/klang/pianissimo -c
ostt model params berget/klang/pianissimoOSTT waits until recording stops, converts the saved audio with ffmpeg to 24 kHz mono PCM16, sends it in chunks to wss://api.berget.ai/v1/realtime?intent=transcription, and commits one turn. It returns only the final transcript. Microphone audio is not sent during recording, and partial transcripts are not displayed. Berget documents this file-over-WebSocket workflow; Pianissimo is not sent to the Whisper /v1/audio/transcriptions endpoint.
[berget."klang/pianissimo".params]
language = "sv"
chunk_seconds = 3| Param | Type | Description |
|---|---|---|
language | string | Defaults to sv. Sent as Berget's languages array; Pianissimo is a Swedish model. |
chunk_seconds | number | Positive, finite server-side target segment length in seconds; defaults to 3. Smaller segments can reduce context. This does not control recording duration or enable live output. |
For example, use --param chunk_seconds=2 to change the server segmentation. This is distinct from Whisper's chunk_size parameter.
Berget states that Pianissimo ignores prompts, temperature, and keywords. OSTT does not send saved ostt keyword terms to this model and rejects unsupported params, including hotwords, prompt, temperature, alignment, diarization, and response-format options. Keep Whisper-specific settings under [berget."KBLab/kb-whisper-large".params] or another Whisper model's section, rather than [berget.params], because provider-wide settings also apply to Pianissimo.
Requires a Berget API key, network access to the WebSocket endpoint, and ffmpeg. This is cloud transcription, not local inference. Connection errors, server errors, and a connection closing before a final transcript are reported as failures rather than returning partial text.
Whisper params
[berget."KBLab/kb-whisper-large".params]
language = "sv"
hotwords = ["OSTT", "KBLab", "Berget"]
prompt = "Swedish technical dictation."
temperature = 0.0
response_format = "verbose_json"
align = true
diarize = trueostt transcribe meeting.mp3 -m berget/KBLab/kb-whisper-large --param language=sv --param hotwords=OSTT,KBLab --param align=true
ostt model params berget/KBLab/kb-whisper-large --format jsonOSTT always returns plain transcript text. verbose_json, word alignment, and diarization params are supported because Berget responses still include a top-level text field; metadata is not emitted in command output.
Whisper uploads use a MIME type selected from the file extension. Timestamp granularities are sent as a CSV field. Requests have a 10-second connection timeout and a 30-minute total timeout covering upload, processing, and response reading.
| Param | Type | Description |
|---|---|---|
language | string | Optional language hint, such as sv or no. |
hotwords | string list | Berget keyword boosting terms. Saved ostt keyword terms are used as fallback only when hotwords is not set. |
prompt | string | Whisper-compatible context prompt. Saved ostt keyword terms are used as fallback only when prompt is not set. |
temperature | number | Sampling temperature, 0.0 to 1.0. |
response_format | string | Supported values in OSTT: json, verbose_json. Berget also documents text, srt, and vtt, but OSTT does not expose them because this provider path parses JSON. |
timestamp_granularities | string list | Supported values: word, segment. |
align | boolean | Enable word-level timestamp alignment. Berget documents this as adding word start/end timestamps and confidence scores. |
diarize | boolean | Enable speaker diarization with automatic speaker labels. |
speaker_embeddings | boolean | Enable speaker embeddings. |
chunk_size | integer | Chunk size in seconds, 1 to 60. |
batch_size | integer | Processing batch size, 1 to 32. |
Whisper file-upload limits
Berget documents supported upload formats mp3, mp4, mpeg, mpga, m4a, wav, and webm, with a maximum file size of 100 MB and a maximum processing time of 30 minutes per request. These limits describe the Whisper file-upload route, not Pianissimo's WebSocket transport. OSTT does not expose a stream param for either route.