Audio persistence

SpeechRouter’s core promise is that audio passes straight through to the provider and nothing else ever sees it — that’s true for the hosted api.speechrouter.ai unconditionally, and it’s the default for self-hosted gateways too. If you have a genuine reason to keep a copy of session audio — compliance, QA review, dispute resolution — you can opt in on your own self-hosted deployment. It is never available on the hosted service.

Enabling it

SPEECHROUTER_AUDIO_SINK=s3
SPEECHROUTER_AUDIO_SINK_BUCKET=your-bucket
SPEECHROUTER_AUDIO_SINK_PREFIX=recordings        # optional

Credentials and region default to the same SPEECHROUTER_AWS_* variables used for the AWS Transcribe provider. Set the _AUDIO_SINK_ variants only if recordings should go to a different account, bucket, or region:

Variable	Default	Meaning
SPEECHROUTER_AUDIO_SINK	none	none | s3
SPEECHROUTER_AUDIO_SINK_BUCKET	—	Required when s3
SPEECHROUTER_AUDIO_SINK_PREFIX	—	Optional key prefix, e.g. recordings
SPEECHROUTER_AUDIO_SINK_REGION	SPEECHROUTER_AWS_REGION	Override if different from Transcribe’s region
SPEECHROUTER_AUDIO_SINK_ACCESS_KEY	SPEECHROUTER_AWS_ACCESS_KEY_ID	Override for a separate account
SPEECHROUTER_AUDIO_SINK_SECRET_KEY	SPEECHROUTER_AWS_SECRET_ACCESS_KEY	Override for a separate account

Your IAM policy needs s3:PutObject (and s3:AbortMultipartUpload for cleanup on error paths) on the bucket.

What gets stored

One object per session, at {prefix}/{session_id}.raw — the raw audio bytes exactly as received, in whatever encoding and sample rate the session declared. Raw PCM isn’t self-describing on its own; a consumer that needs a playable file prepends a WAV header using the session’s known encoding/rate.

How it’s written

Streamed via real S3 multipart upload as audio arrives — buffered only up to S3’s 5 MB minimum part size per in-flight part, never a whole session. A long session or many concurrent sessions never spikes gateway memory. Sessions under 5 MB total skip multipart entirely and land as a single PutObject.

A sink failure never breaks transcription — write errors are logged, not raised into the session, and a failed multipart upload is aborted server-side rather than left as an orphaned (and billed) incomplete upload.

Building your own sink

SPEECHROUTER_AUDIO_SINK=s3 ships in the box, but the hook itself is a two-method interface:

class AudioSink(Protocol):
    async def on_chunk(self, session_id: str, chunk: bytes) -> None: ...
    async def on_session_end(self, session_id: str) -> None: ...

Anything satisfying that — write to local disk, a different object store, a message queue — plugs in the same way; see speechrouter_gateway/audio/sink.py and s3_sink.py for the reference shape.