How to turn calls into text

Steps

1. Get the call as a file

   Use the recording your call app saves, or record the call with the phone or a screen recorder. Tell the people on the call that you record it: rules differ by country.

2. Shrink it

   Speech does not need a heavy file. One mono track is enough, and a small file uploads and runs faster. In this command, mono at 32 kbit/s is about 14 MB an hour; the flags are standard ffmpeg options (-ac channels, -ar sample rate, -b:a audio bitrate).

   $ ffmpeg -i call.m4a -ac 1 -ar 16000 -b:a 32k call.mp3

   The Uno gateway takes ogg, mp3, wav, mp4, m4a, webm and flac, up to 25 MB. OpenAI's own transcription endpoint takes files up to 25 MB, and advises compressing or splitting longer audio, not mid-sentence.

3. Pick a route

   An Uno computer has no GPU. Run the small Whisper models on its CPU, or use the gateway.

   	Whisper	faster-whisper	Uno gateway
   Runs on	Your machine	Your machine	Uno
   Install	pip install -U openai-whisper and ffmpeg	pip install faster-whisper	Nothing
   Needs	PyTorch; turbo wants about 6 GB of video memory	on a CPU it needs nothing beyond Python; on an NVIDIA GPU it needs cuBLAS and cuDNN 9 for CUDA 12	An Uno API key
   Cost	Free, MIT, for the code and the model weights	Free, MIT	Part of Uno AI

4. Route 1: Whisper

   Whisper also needs the command-line tool ffmpeg on the system. Install both, then run one command. Whisper writes the text and subtitle files next to the audio.

   $ pip install -U openai-whisper
   $ whisper audio.mp3 --model turbo
   # other models: tiny, base, small, medium, large; whisper --help lists the options

   The models trade speed for accuracy. tiny: 39 million parameters, about 1 GB of video memory, about 10 times the speed of large. small: 244 million parameters, about 2 GB of video memory, about 4 times the speed of large. turbo: 809 million parameters, about 6 GB of video memory, about 8 times the speed of large. large: 1550 million parameters, about 10 GB of video memory.

5. Route 2: faster-whisper

   It is up to 4 times faster than openai/whisper for the same accuracy, with less memory, and it runs on a CPU. It supports compute types float16, int8_float16 and int8. The int8 type is the light one for a CPU.

   $ pip install faster-whisper

   from faster_whisper import WhisperModel
   model = WhisperModel("small", device="cpu", compute_type="int8")
   segments, info = model.transcribe("call.mp3")
   for s in segments:
       print("[%.1fs -> %.1fs] %s" % (s.start, s.end, s.text))

6. Route 3: the Uno gateway

   Send one request: POST /v1/audio/transcriptions, multipart, OpenAI-compatible: file, model, language, response_format. It follows OpenAI's format, so a client written for that format can point here.

   $ curl https://api.getuno.xyz/v1/audio/transcriptions \
       -H "Authorization: Bearer $UNO_API_KEY" \
       -F [email protected] -F model=whisper-large-v3 -F response_format=text

   The reply format is json, text or verbose_json. Take verbose_json when you want timestamps for each segment. Today: Whisper large v3.

7. Turn the text into notes

   A transcript is long. Give it to an agent and ask for what you need: a summary, decisions, tasks with names and dates.

Check that it works

Open the text and read the first minute against the audio. Names, numbers and product words are where it slips. Then look at the end: if the text stops early, the file was cut, so shrink it again or split it.

$ head -c 600 call.txt
$ tail -c 600 call.txt

What it does not do

• Whisper writes text with timestamps; it does not say who is speaking. For "who said what", add a speaker-labelling tool or ask the agent to mark turns from the context.
• Accuracy depends on the language and the audio. Whisper's own README says its performance varies widely by language. Noise, crosstalk and thick accents lower it.
• One file up to 25 MB on the Uno gateway. Split a longer recording into parts and join the texts.
• This guide covers recordings and dictation. It does not join a live call for you.

Common problems