speaker-verification

I wanted a way to grade class participation fairly. Professors do it from memory, which biases toward whoever talks loudest. If you enroll each student against a 30–45 s voice clip, then feed in a lecture recording, you can in theory get a per-student speaking-time breakdown instead of a vibes-based grade.

This repo gets you most of the way there. It wraps the Azure Speaker Recognition text-independent API: upload enrollment audio, upload a verification clip, get back which enrolled voices are present and at what confidence. The piece that's not done yet is time-aligned segmentation (splitting the lecture into short chunks so you can sum per-speaker minutes). Easy to add on top of what's here.

No voice samples shipped. Biometric data shouldn't live in a public repo. Drop your own enrollment clips into audio/enrollment/, verification clips into audio/verification/ (both gitignored), and you can run it.

How it works

  audio/enrollment/*.m4a            audio/verification/*.wav
         │                                   │
         ▼                                   ▼
    convert  (pydub → 16 kHz mono wav)   convert
         │                                   │
         ▼                                   │
    enroll  ── POST profile  ──┐             │
    enroll  ── upload wav  ────┤             │
                              Azure          │
                               │             │
                               ▼             ▼
                           profiles.csv ──► verify
                                              │
                                              ▼
                                 verification_results.csv
                                 (name, score, audio_length)

• enroll creates one Azure profile per person, uploads their clip, and saves (name, profile_id, status, audio_length_s, speech_length_s) to data/profiles.csv.
• verify takes a folder of clips, queries each enrolled profile for a match, and keeps only Accept results with a score above Azure's threshold.

Setup

Requires Python 3.10+ and ffmpeg on your path (pydub depends on it).

git clone https://github.com/tomvlt1/speaker-verification.git
cd speaker-verification
python3 -m venv venv
source venv/bin/activate
pip install -e .
cp .env.example .env           # then fill in your Azure key + region

Get an Azure Speech key: portal.azure.com → create a Speech resource → Keys and Endpoint. The free tier is fine for testing.

Usage

Drop enrollment audio (one file per speaker) into audio/enrollment/, verification audio into audio/verification/. Filename stem becomes the speaker name (trailing _p is stripped).

# 1. normalize m4a/mp3/... → 16 kHz mono wav, then write an enrollment csv
python -m speaker_verification convert \
  --audio-dir audio/enrollment \
  --build-csv enrollment_data.csv

# 2. enroll each speaker against Azure
python -m speaker_verification enroll

# 3. verify a folder of new recordings against the enrolled profiles
python -m speaker_verification convert --audio-dir audio/verification
python -m speaker_verification verify

Results land in data/verification_results.csv.

# nuke all enrolled profiles on the Azure subscription (with a confirmation prompt)
python -m speaker_verification reset
python -m speaker_verification reset --yes   # skip prompt

Output format

data/profiles.csv:

name	profile_id	status	audio_length_s	speech_length_s
alice	5e2a…	Enrolled	37.9	34.9

data/verification_results.csv:

audio_file	audio_length_s	name	profile_id	score
lecture_01.wav	823.4	alice	5e2a…	0.87
lecture_01.wav	823.4	bob	88e7…	0.77

Multiple accept rows per file are possible. Azure verifies independently against each profile, so in a lecture with four enrolled students you can get four matches on the same clip.

Known limitations

• Azure subscription required. This is a thin client; the actual model is Microsoft's. Free tier is rate-limited (20 transactions per second, 10,000 per month at time of writing).
• Text-independent verification needs ≥20 s of speech per enrollment for reliable scores. Azure returns Enrolling rather than Enrolled if you're below that.
• No per-speaker speaking time yet. verify tells you which enrolled speakers appear in a clip but not for how long. To get that, you'd need to split the input into short segments and run verify on each (easy to add, happy to accept PRs).
• Binary Accept/Reject. Azure returns a confidence score 0–1, but there's no timestamp alignment; it's a global judgment on the whole clip.
• Privacy. Enrollment audio contains biometric data. Don't commit it; the .gitignore here already protects audio/**/*.

Layout

src/speaker_verification/
  __main__.py   # argparse dispatcher
  api.py        # single HTTP helper for Azure calls
  config.py     # env vars (.env loader) + project paths
  convert.py    # ffmpeg → 16 kHz mono wav + enrollment csv builder
  enroll.py     # create profiles + upload enrollment clips
  verify.py     # verify unknown audio against all profiles
  reset.py      # delete every profile on this subscription
audio/
  enrollment/   # your clips go here (gitignored)
  verification/ # your clips go here (gitignored)
data/           # generated csvs land here (gitignored)