pytest-audioeval

Pytest plugin for STT/TTS integration testing. Built on the httpx ecosystem (httpx, httpx-ws, httpx-sse) with built-in metrics, embedded ground-truth audio samples, and chainable assertions.

Features

• STT via WebSocket — audioeval.stt.ws() streams audio, collects transcription
• TTS via HTTP — audioeval.tts.post() batch, .stream() chunked, .sse() Server-Sent Events
• Text metrics — WER, CER, substitutions, insertions, deletions (via jiwer)
• Audio metrics — PESQ MOS 1–5 scale (via pesq)
• Embedded samples — ground-truth audio + reference text pairs, multi-language ready
• Chainable assertions — result.compute_metrics(ref).assert_quality(max_wer=0.2)
• CLI thresholds — --audioeval-wer, --audioeval-cer, --audioeval-mos

Install

uv add pytest-audioeval

Quick Start

STT — WebSocket

import asyncio
import uuid
import orjson as json
from pytest_audioeval.client import AudioEval


async def test_user_stt_ws(audioeval: AudioEval) -> None:
    sample = audioeval.samples.en_hello_world

    async with audioeval.stt.ws(sample=sample) as session:
        config = json.dumps(
            {"uid": str(uuid.uuid4()), "language": "en", "task": "transcribe",
             "model": "large-v3-turbo", "use_vad": True}
        ).decode()
        await session.send_text(config)

        ready = await session.receive_text()
        assert "SERVER_READY" in ready

        await session.send_sample(sample, chunk_ms=200)
        await asyncio.sleep(2)
        await session.send_text("END_OF_AUDIO")

        # Collect transcription segments...

TTS — Batch POST

import io
import soundfile as sf
from pytest_audioeval.client import AudioEval


async def test_user_tts_batch(audioeval: AudioEval) -> None:
    response = await audioeval.tts.post(
        json={"input": "Hello world.", "model": "kokoro",
              "voice": "af_heart", "response_format": "wav", "stream": False},
    )
    data, rate = sf.read(io.BytesIO(response.content), dtype="float32")
    assert rate == 24_000
    assert len(data) > 0

TTS — Chunked Streaming

async def test_user_tts_streaming(audioeval: AudioEval) -> None:
    chunks = []
    async with audioeval.tts.stream(json={"input": "Hello.", ...}) as response:
        async for chunk in response.aiter_bytes():
            chunks.append(chunk)
    assert len(chunks) > 0

TTS — Server-Sent Events

async def test_user_tts_sse(audioeval: AudioEval) -> None:
    async with audioeval.tts.sse(json={"input": "Hello.", ...}) as event_source:
        async for sse in event_source.aiter_sse():
            print(sse.data)

Text Metrics

from pytest_audioeval.metrics.text import TextMetrics


async def test_user_metrics_text() -> None:
    metrics = TextMetrics.compute(
        reference="the quick brown fox jumps over the lazy dog",
        hypothesis="the quick brown fox jumps over the lazy dock",
    )
    assert metrics.wer < 0.15
    assert metrics.substitutions == 1

STT Result — Chainable Assertions

from pytest_audioeval.stt import STTResult


async def test_user_stt_result() -> None:
    result = STTResult(hypothesis_text="Hello world.")
    result.compute_metrics("Hello world.")
    result.assert_quality(max_wer=0.2, max_cer=0.15)

Sample Registry

from pytest_audioeval.samples.registry import SampleLang


async def test_user_samples_browse(audioeval: AudioEval) -> None:
    # All samples
    assert len(audioeval.samples) >= 3

    # Filter by language
    en_samples = audioeval.samples.by_lang(SampleLang.EN)

    # Attribute access: {lang}_{name}
    sample = audioeval.samples.en_hello_world
    assert sample.reference_text == "Hello world."

    # Audio access
    audio_f32 = sample.audio_numpy()        # numpy float32 array
    audio_raw = sample.audio_bytes()         # raw bytes
    chunks = sample.chunks(chunk_ms=200)     # chunked for streaming

CLI Thresholds

async def test_user_thresholds(audioeval_thresholds: dict[str, float]) -> None:
    assert audioeval_thresholds["max_wer"] == 0.2
    assert audioeval_thresholds["max_cer"] == 0.15
    assert audioeval_thresholds["min_mos"] == 3.0

CLI Options

pytest --stt-url=ws://localhost:45120 --tts-url=http://localhost:45130/v1/audio/speech
pytest --audioeval-wer=0.15 --audioeval-cer=0.10 --audioeval-mos=3.5

Fixtures

Architecture

src/pytest_audioeval/
├── plugin.py              # pytest entry point (fixtures, CLI options)
├── client.py              # AudioEval facade
├── stt.py                 # STTClient (httpx-ws), STTSession, STTResult
├── tts.py                 # TTSClient (httpx + httpx-sse)
├── metrics/
│   ├── text.py            # TextMetrics — WER, CER via jiwer
│   └── audio.py           # AudioMetrics — PESQ MOS via pesq
└── samples/
    ├── registry.py        # SampleRegistry + AudioSample + SampleLang
    └── audio/en/          # Embedded ground-truth WAV + TXT pairs

Clients

Metrics

Samples

Embedded ground-truth audio with reference transcriptions:

samples/audio/
└── en/                    # English (16kHz, float32)
    ├── hello_world.wav    # "Hello world."
    ├── quick_brown_fox.wav
    └── counting.wav       # "One, two, three, four, five."

Access: audioeval.samples.en_hello_world, audioeval.samples.en_counting, etc.

Infrastructure

Integration tests require GPU-accelerated TTS/STT services:

make infra-up       # Start TTS (Kokoro) + STT (WhisperLive)
make infra-status   # Check health
make infra-logs     # View logs
make infra-down     # Stop services

Development

make install            # uv sync --dev
make lint               # ruff check + format
make test-unit          # unit tests (no services)
make test-integration   # integration tests (requires services)
make coverage           # coverage report (>90%)

Requirements

• Python >= 3.13
• NVIDIA GPU + Docker with nvidia-container-toolkit (for integration tests)

License

MIT