gigaam-api

HTTP service that turns a stereo call recording into a speaker-separated transcript with emotion scores. Built on GigaAM (ASR and emotion models) and Silero VAD for speech segmentation.

The stereo assumption replaces heavy diarisation: the left channel is treated as the client, the right channel as the operator. Each channel is sliced by VAD, every speech fragment is transcribed, and emotion probabilities are estimated per fragment.

API

POST /analyze

multipart/form-data with audio_file; optional query parameters:

Parameter	Default	Meaning
min_silence_duration_ms	500	pause that splits a phrase
speech_pad_ms	100	padding around detected speech

Returns a list of utterances sorted by start time:

[
  {
    "text": "здравствуйте, меня зовут анна",
    "start": 1.24,
    "end": 3.9,
    "speaker": "OPERATOR",
    "emotions": {"neutral": 0.81, "positive": 0.12, "...": 0.0},
    "dominant_emotion": "neutral"
  }
]

Mono input is rejected with 400: without two channels the speaker labels would be a guess.

Behaviour and limits

• audio is resampled to 16 kHz and processed per channel;
• fragments longer than 25 seconds are split, because GigaAM accepts up to 25 seconds;
• torch.load is patched for PyTorch 2.6+, where the default weights_only=True breaks the model checkpoints;
• models load once at startup and stay in memory — the first request is the slowest;
• CPU works, CUDA is used automatically when available (GIGAAM_DEVICE).

Configuration

Variable	Default	Meaning
GIGAAM_MODEL	v3_e2e_rnnt	ASR model; v3_e2e_rnnt is the phone-call default
GIGAAM_DEVICE	cuda if available, else cpu	inference device

Run

docker build -t gigaam-api .
docker run --rm -p 9000:9000 gigaam-api

curl -F "audio_file=@call.wav" \
  "http://127.0.0.1:9000/analyze?min_silence_duration_ms=500&speech_pad_ms=100"

Model checkpoints are downloaded on first start; mount a cache volume at /root/.cache to keep them between runs.

License

MIT for this service. GigaAM models and code are MIT (salute-developers), Silero VAD is MIT as well.