Local speech-to-text¶
Outcome: POST /v1/audio/transcriptions and POST /v1/audio/translations stay on the machine. A cloud upload happens only when you opt in.
Config¶
asr.base_url is an OpenAI-compatible API root (it includes /v1), the same shape as frontier.base_url. Point it at vLLM, whisper.cpp's OpenAI server, or any pool member that already serves POST /v1/audio/transcriptions and POST /v1/audio/translations. No Whisper or ffmpeg package is installed with daari.
asr:
base_url: http://127.0.0.1:8000/v1
model: "" # optional; when set, replaces the client model
frontier_fallback: false # default — never upload audio implicitly
Leave base_url empty and the route returns 501. daari doctor stays quiet in that case. It warns when asr.base_url is set but GET {base}/models is unreachable, and when frontier_fallback is true while frontier is disabled or no API key resolves. frontier_fallback stays off so existing installs do not start sending meetings to a hosted API. Set it to true only when frontier.enabled is true and a frontier key is configured; daari then forwards the transcription to eligible frontier slots in pool order (failover on transient failure). Virtual-key no_frontier blocks the upload (403); region_pin keeps audio on matching-region slots only (400 when none match).
Helm fleets can set asr.baseUrl, which mounts DAARI_ASR__BASE_URL. Optional chart value asr.model mounts DAARI_ASR__MODEL so the pod always presents the on-box whisper id. asr.frontierFallback defaults off and omits env; set it to true to mount DAARI_ASR__FRONTIER_FALLBACK so an empty base URL may forward one transcription to frontier (see Capacity and Helm).
Chat input_audio blocks¶
OpenAI chat completions and Responses message content may include
{"type": "input_audio", "input_audio": {"data": "<base64>", "format": "wav"|"mp3"}}
parts. daari keeps those on the internal message and rebuilds them on the L6 OpenAI
payload so frontier never sees a silently stripped turn. When asr.base_url is set,
daari also POSTs each clip to the local ASR endpoint and appends the transcript to the
message text before local tiers run — voice-agent turns can stay on-box. With ASR unset,
local tiers see only the text caption (if any); the raw audio still rides to frontier on
escalation.
Request¶
OpenAI multipart form: file, model (required unless asr.model is set), optional language (transcriptions only), prompt, and response_format=json. Translations omit language and forward to {base}/audio/translations. The JSON body includes text. Leave base_url empty and either route returns 501 without uploading. response_format other than json is 400.
Auth, virtual-key request budgets, and rate limits apply the same way as chat completions. A key or team model allowlist is checked against the resolved model (asr.model when set, otherwise the form model) and returns 403 model_not_allowed before any upstream call. Unset allowlists stay unrestricted. A successful transcription counts as one request.
Verify¶
With asr.base_url set, a short clip returns { "text": "..." }. With nothing configured, the same call returns 501 and the file is not uploaded.