Configuration¶
The backend is configured with environment variables.
Variable |
Required |
Description |
|---|---|---|
|
Default provider |
Alibaba Cloud DashScope key for Qwen3-ASR-Flash, Qwen Omni Realtime and Qwen-Audio ASR. The default provider is |
|
No |
Model Studio workspace id. Needed only by |
|
No |
Comma-separated |
|
No |
Groq API key for gpt-oss-120b post-processing; enables the |
|
No |
Default configuration id (default: derived from |
|
No |
Comma-separated configurations clients may select (default: every configuration with credentials) |
|
No |
Provider for the derived default: |
|
No |
Model for the derived default (default: the provider’s own default model) |
|
No |
BytePlus key; enables the |
|
No |
BytePlus language hint, e.g. |
|
No |
BytePlus resource id (default: |
|
No |
BytePlus endpoint base, without the mode path segment |
|
No |
OpenAI key; enables the |
|
No |
xAI key; enables the |
|
No |
ElevenLabs key; enables the |
|
No |
iApp key; enables the |
|
No |
OpenRouter key; enables the |
|
No |
Meta key; enables the |
|
No |
Paxa Labs key; enables the |
|
No |
|
|
No |
Google Gemini API key; enables the |
|
No |
Vercel AI Gateway key; enables the |
|
No |
OIDC provider discovery URL (alternative to API keys) |
|
No |
OIDC client ID (default: |
|
No |
Expected token audience (default: same as client ID) |
|
No |
HMAC secret for session tokens (default: |
|
No |
Address clients should use for pairing links (default: the address the request came in on) |
|
No |
Endpoint to POST completed transcriptions to |
|
No |
HTTP port (default: |
|
No |
How long a recording socket may go without audio before it is closed and its vendor connection released (default: |
|
No |
How long an unused Qwen Omni connection is kept open after a recording, for the next one from the same client (default: |
|
No |
Audio a kept Qwen Omni connection may carry across recordings before it is retired (default: |
Eval storage¶
Where eval votes and eval-sets are saved. The backend only signs upload URLs; the browser uploads directly to the bucket, so recordings never pass through the server and no bucket credential ever reaches a browser. Setting EVAL_STORAGE_BUCKET turns the feature on; without it, evals still run and winners still apply, they are just not recorded.
Variable |
Required |
Description |
|---|---|---|
|
No |
S3-compatible bucket for votes and eval-sets. Enables eval storage |
|
With bucket |
Access key id used to sign upload URLs |
|
With bucket |
Secret access key used to sign upload URLs |
|
No |
Bucket region (default: |
|
No |
Custom S3-compatible endpoint, e.g. MinIO or R2 (default: AWS) |
|
No |
Set to |
|
No |
Key prefix placed before |
Objects are keyed by kind, then UTC date, then time and message id:
votes/2026/07/16/2026-07-16T09-30-00-123Z-<messageId>.json
eval-sets/2026/07/16/2026-07-16T09-30-00-123Z-<messageId>.json
A vote ({"type":"vote",…}) is written on every winner pick and carries the winning configuration id, the full candidate list, per-candidate cost/latency/errors, and the audio’s duration — but no audio and no transcripts, so it is safe to store unconditionally. An eval-set ({"type":"eval-set",…}) is written only when the user ticks save-for-eval, and adds the recording as a base64 WAV plus every candidate’s transcript. A pick’s vote and eval-set share a key suffix, so one finds the other.