# ASR: one capability, two providers Requires the **1.9.3rc23 preview** and FFmpeg, including `ffprobe`, even for short files. Stable 1.9.2 does not include `co asr`. The [package](https://pypi.org/project/connectonion/1.9.3rc23/) and [GitHub prerelease](https://github.com/openonion/connectonion/releases/tag/v1.9.3rc23) are published. Install this exact preview without changing the stable recommendation: ```bash python -m pip install --upgrade 'connectonion==1.9.3rc23' brew install ffmpeg ``` The FFmpeg installation example is for macOS. On other platforms install `ffmpeg` and `ffprobe` with your package manager and put both on PATH. ```bash co asr recording.m4a --language zh co asr recording.m4a --json co env set ASR_PROVIDER connectonion co asr recording.m4a co asr recording.m4a --provider codex co asr recording.m4a --resume JOB_ID ``` Replace `recording.m4a` with your existing audio file. The default provider is `codex`; `ASR_PROVIDER` in co's existing selected env file changes it, and `--provider` overrides it once. No separate config file or server is needed. `co --env-file /path/to/settings.env asr recording.m4a` selects a specific co configuration through the existing CLI. Python callers configure their process environment or pass the provider explicitly. ```python from connectonion import asr result = asr("recording.m4a", provider="codex", language="zh") print(result["text"]) ``` Both providers upload audio. Codex uses your existing file-backed ChatGPT login under `$CODEX_HOME/auth.json` or `~/.codex/auth.json`. It does not refresh or modify that file, ask you to paste a token, or send the token to OpenOnion. Keychain-only Codex credentials and API-key-only login are not supported by this adapter. If file-backed login is missing or expired, run `codex login` and check your Codex credential-store configuration. Installing Codex alone does not imply that ASR access or unlimited subscription usage is available. ConnectOnion uses the existing `transcribe()` API, co authentication and credits. `ASR_MODEL` may select a `co/` model, otherwise the SDK default applies. This setting applies only to the ConnectOnion provider. Existing `transcribe()` calls retain their original behavior. No provider fallback or automatic retry occurs. The original file is preserved. Text goes to stdout and destination/progress to stderr. Short direct requests return `{ok, provider, text}` with `--json`; long recordings also return a `job_id` and raw `segments` with source-window start/end offsets. Those offsets are not word-level timestamps. Errors omit upstream response bodies. ## Long meetings and interrupted work There is no 50 MiB input-file limit. Install FFmpeg (`ffprobe` and `ffmpeg`) for local inspection and conversion. For example on macOS, `brew install ffmpeg`. WAV, MP3, M4A/MP4, FLAC, Ogg and WebM can go directly to the provider when short; other decodable formats are converted locally. There is no upload just to discover duration. Long audio is processed one eight-minute window at a time with two seconds of boundary overlap, converted to 16kHz mono PCM. A three-hour meeting becomes 23 requests rather than a rejected whole file. Only the prepared individual request has an internal size budget; the original can be larger. The upstream project's eleven-minute recommendation is not an official Codex limit. This window size is a conservative implementation choice, not a service guarantee. Completed raw transcripts are atomically checkpointed in private, owner-only directories at `~/.co/asr/jobs/JOB_ID/state.json` (or under `AGENT_CONFIG_PATH`). These local transcript files are plaintext; original audio is not copied into the checkpoint. Temporary converted windows are removed after each part. Delete the job directory when you no longer need its saved transcripts/recovery. On interruption, `--json` reports `job_id`, `completed_segments` and `partial_text`. Resume with the same original audio and the returned ID: ```bash co asr recording.m4a --resume JOB_ID --json ``` Resume verifies the original's streamed SHA-256, keeps the recorded provider, model and language even if default configuration changed, and skips already checkpointed parts. A second process cannot run the same job simultaneously. An uncertain request or a failed checkpoint write may require submitting that one part again; remote exactly-once billing is not promised. No other provider is invoked to conceal a failure. The returned text concatenates raw parts in order and preserves repeated words; two-second boundary overlap may repeat speech. Text similarity alone cannot prove that a repetition is accidental, so this layer never silently deletes it. Raw part transcripts remain separate; any later AI cleanup is another version, not an overwrite. Empty text can represent silence. No speaker labels, word timestamps or automatic REM ingestion. Network timeout is 15 seconds to connect and 300 seconds of network inactivity per request. Keep originals to check omissions or recognition errors. ## Protocol source and experimental status The Codex provider is an independent Python implementation informed by [Wangnov/codex-asr](https://github.com/Wangnov/codex-asr), an MIT-licensed Rust project. It uses `https://chatgpt.com/backend-api/transcribe` with multipart audio, bearer authentication, an optional account routing header and a JSON text response. This is the unofficial Codex Desktop dictation endpoint, not the public OpenAI transcription API or Codex app-server realtime. Availability and subscription entitlement are not guaranteed and may change. No external `codex-asr` binary, Rust runtime, token broker or REST wrapper is installed. Protocol compatibility tests use synthetic credentials and HTTP; those tests do not establish live account access or recognition quality.