Audio formats

Audio you send

/v1/transcribe, /v1/translate and /v1/converse read the container from the file’s first bytes, so the file name and content type do not matter.

Container	Codecs	Where it comes from
OGG	Opus	WhatsApp and Telegram voice notes, Android
M4A, MP4, 3GP, AAC	AAC	iPhone voice memos, many Android recorders
MP3	MP3	Radio clips, downloads
WAV	PCM	Computers, studio recorders
WebM	Opus	Browsers (`MediaRecorder`)
FLAC	FLAC	Archives

Up to 25 MB and 300 seconds per request. Mono or stereo, any sample rate.

Two ways to send it

Multipart form
Raw body

The usual way. The file goes in a field named file, the options beside it.

curl https://api.namuai-lab.com/v1/transcribe \
  -H "Authorization: Bearer $NAMU_API_KEY" \
  -F file=@note.ogg -F language=ha

The bytes are the whole body; options go in the query string. Handy when the audio is already in memory.

curl "https://api.namuai-lab.com/v1/transcribe?language=ha" \
  -H "Authorization: Bearer $NAMU_API_KEY" \
  --data-binary @note.ogg

Audio you get back

/v1/speech returns what you ask for with format:

`format`	Content type	Details	Size of 10 seconds
`wav` (default)	`audio/wav`	24 kHz, 16-bit PCM, mono	about 480 KB
`mp3`	`audio/mpeg`	64 kbit/s, mono	about 80 KB
`opus`	`audio/ogg`	Ogg Opus, 48 kHz, 32 kbit/s, mono	about 40 KB
`json`	`application/json`	base64 WAV in `audio`	about 640 KB

/v1/translate (with include_audio) and /v1/converse return the spoken answer as base64 WAV in the JSON.

For people on mobile data, ask for opus: it is a twelfth of the size of WAV and plays in Chrome, Firefox, Android and WhatsApp. For Safari on older iPhones, use mp3.

Long recordings

A request takes up to 5 minutes of audio. For longer recordings, cut them into pieces at silences and send the pieces in order:

# Cut into 4-minute pieces with ffmpeg, without re-encoding
ffmpeg -i meeting.m4a -f segment -segment_time 240 -c copy piece-%03d.m4a

Then transcribe piece-000.m4a, piece-001.m4a, … and join the texts. A cut in the middle of a word can lose that word; cutting at a pause (for example with ffmpeg’s silencedetect) avoids it.

Text to speech

Build

Audio you send

Two ways to send it

Audio you get back

Long recordings

​Audio you send

​Two ways to send it

​Audio you get back

​Long recordings

Audio you send

Two ways to send it

Audio you get back

Long recordings