What Nanoru does
Nanoru takes recordings and live streams in English and returns transcripts in which every line has a speaker.
What goes in
- Audio and video files
- You can upload files in the dashboard or send them over the API.
- An archive
- Connect a storage bucket and Nanoru processes what is in it, along with anything you add later.
- A podcast feed
- Add a feed and Nanoru processes each new episode as it is published. You choose how far back it goes.
- A live stream
- Nanoru transcribes a live stream as it plays.
What comes out
- A dashboard
- Your team can read and play back every recording in the dashboard, and see who speaks in it.
- JSON over an API
- The API returns each line with its speaker and word timings, and a webhook tells you when a result changes.
- Captions
- You can export SRT and VTT captions with the speakers’ names in them.
How a voice gets its name
Nanoru groups the speech in a recording by voice, then compares each voice with the voiceprints your organisation already has. A match gets that person’s name.
A voice with no match is an unknown speaker, with a label such as Unknown #8. The same voice keeps that label in every recording, so you can see how often an unknown speaker appears.
When someone introduces a guest by name, Nanoru suggests that name for the guest’s voice. You confirm it, or type another, and every recording that voice is in is updated.
Nanoru builds each voiceprint from lines you have confirmed, so nobody has to record a sample.
The Friday Panel, episode 213, voices
1Unknown #8in 3 episodes
Suggested: Maria Lopez, from “With us this week is the chef Maria Lopez” at 0:59.
- “Thanks for having me back.”
- “Then I’d like a mug with my name on it.”
- Enter yes, Maria Lopez
- N type another name
- 2 same person as Sam Okafor
- 0 not a speaker
2Sam Okafor
3Alex Reid
What each line tells you
The API returns the same thing the dashboard shows. Each line has:
- A speaker
- This is a named person, or an unknown speaker with a label.
- Confidence
- This is how sure Nanoru is of the speaker, from 0 to 1. It appears once Nanoru has been calibrated on a sample of your recordings that a person has checked.
- Overlapping speech
- Nanoru marks a line when someone else was talking at the same time.
- Not speech
- Nanoru finds the adverts, intros and jingles that repeat between episodes. A reviewer can mark anything else, such as music.
- Language
- Nanoru transcribes English. It marks speech in another language with the name of the language and does not transcribe it.
{
"speakers": [
{ "id": "0197a3c2-51d0-7b1e-9c04-2f6d8a1e5b90",
"kind": "named", "name": "Sam Okafor",
"label": "Sam Okafor" },
{ "id": "0197a3c2-6e44-7f02-b8a1-91c3d7e02a4c",
"kind": "unknown", "label": "Unknown #8" }
],
"turns": [
{ "start_s": 11.2, "end_s": 14.9,
"speaker_id": "0197a3c2-51d0-7b1e-9c04-2f6d8a1e5b90",
"confidence": 0.994,
"overtalk": false,
"non_speech": null,
"words": [
{ "w": "Hello", "start_s": 11.2, "end_s": 11.5 }
] }
]
}On a live stream
Words appear a few seconds after they are spoken. A speaker Nanoru has a voiceprint for is named as they talk.
Every few minutes it goes back over what it has heard and corrects words and names. When the stream ends it processes the whole recording as it would a file, and that final version is the one our accuracy figures describe.
Your systems get each new line and each correction as it happens, through the API.
Test it on your own recordings
Send 20 episodes or 20 hours of recordings. We return them with your presenters named, at no charge.