selfhostyourtech/reflector
2
0
Fork 0
mirror of https://github.com/Monadical-SAS/reflector.git synced 2025-12-20 12:19:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: align whisper transcriber api with parakeet (#602)

Browse Source 
* Documents transcriber api

* Update whisper transcriber api to match parakeet

* Update api transcription spec

* Return 400 for unsupported file type

* Add params to api spec

* Update whisper transcriber implementation to match parakeet
This commit is contained in:
Sergey Mankovsky
2025-09-05 10:52:14 +02:00
committed by GitHub
parent dc82f8bb3b
commit 0663700a61
3 changed files with 705 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

194
server/docs/gpu/api-transcription.md Normal file
View File

@@ -0,0 +1,194 @@
## Reflector GPU Transcription API (Specification)
This document defines the Reflector GPU transcription API that all implementations must adhere to. Current implementations include NVIDIA Parakeet (NeMo) and Whisper (faster-whisper), both deployed on Modal.com. The API surface and response shapes are OpenAI/Whisper-compatible, so clients can switch implementations by changing only the base URL.
### Base URL and Authentication
- Example base URLs (Modal web endpoints):
- Parakeet: `https://<account>--reflector-transcriber-parakeet-web.modal.run`
- Whisper: `https://<account>--reflector-transcriber-web.modal.run`
- All endpoints are served under `/v1` and require a Bearer token:
```
Authorization: Bearer <REFLECTOR_GPU_APIKEY>
```
Note: To switch implementations, deploy the desired variant and point `TRANSCRIPT_URL` to its base URL. The API is identical.
### Supported file types
`mp3, mp4, mpeg, mpga, m4a, wav, webm`
### Models and languages
- Parakeet (NVIDIA NeMo): default `nvidia/parakeet-tdt-0.6b-v2`
- Language support: only `en`. Other languages return HTTP 400.
- Whisper (faster-whisper): default `large-v2` (or deployment-specific)
- Language support: multilingual (per Whisper model capabilities).
Note: The `model` parameter is accepted by all implementations for interface parity. Some backends may treat it as informational.
### Endpoints
#### POST /v1/audio/transcriptions
Transcribe one or more uploaded audio files.
Request: multipart/form-data
- `file` (File) — optional. Single file to transcribe.
- `files` (File[]) — optional. One or more files to transcribe.
- `model` (string) — optional. Defaults to the implementation-specific model (see above).
- `language` (string) — optional, defaults to `en`.
- Parakeet: only `en` is accepted; other values return HTTP 400
- Whisper: model-dependent; typically multilingual
- `batch` (boolean) — optional, defaults to `false`.
Notes:
- Provide either `file` or `files`, not both. If neither is provided, HTTP 400.
- `batch` requires `files`; using `batch=true` without `files` returns HTTP 400.
- Response shape for multiple files is the same regardless of `batch`.
- Files sent to this endpoint are processed in a single pass (no VAD/chunking). This is intended for short clips (roughly ≤ 30s; depends on GPU memory/model). For longer audio, prefer `/v1/audio/transcriptions-from-url` which supports VAD-based chunking.
Responses
Single file response:
```json
{
"text": "transcribed text",
"words": [
{ "word": "hello", "start": 0.0, "end": 0.5 },
{ "word": "world", "start": 0.5, "end": 1.0 }
],
"filename": "audio.mp3"
}
```
Multiple files response:
```json
{
"results": [
{"filename": "a1.mp3", "text": "...", "words": [...]},
{"filename": "a2.mp3", "text": "...", "words": [...]}]
}
```
Notes:
- Word objects always include keys: `word`, `start`, `end`.
- Some implementations may include a trailing space in `word` to match Whisper tokenization behavior; clients should trim if needed.
Example curl (single file):
```bash
curl -X POST \
-H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
-F "file=@/path/to/audio.mp3" \
-F "language=en" \
"$BASE_URL/v1/audio/transcriptions"
```
Example curl (multiple files, batch):
```bash
curl -X POST \
-H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
-F "files=@/path/a1.mp3" -F "files=@/path/a2.mp3" \
-F "batch=true" -F "language=en" \
"$BASE_URL/v1/audio/transcriptions"
```
#### POST /v1/audio/transcriptions-from-url
Transcribe a single remote audio file by URL.
Request: application/json
Body parameters:
- `audio_file_url` (string) — required. URL of the audio file to transcribe.
- `model` (string) — optional. Defaults to the implementation-specific model (see above).
- `language` (string) — optional, defaults to `en`. Parakeet only accepts `en`.
- `timestamp_offset` (number) — optional, defaults to `0.0`. Added to each word's `start`/`end` in the response.
```json
{
"audio_file_url": "https://example.com/audio.mp3",
"model": "nvidia/parakeet-tdt-0.6b-v2",
"language": "en",
"timestamp_offset": 0.0
}
```
Response:
```json
{
"text": "transcribed text",
"words": [
{ "word": "hello", "start": 10.0, "end": 10.5 },
{ "word": "world", "start": 10.5, "end": 11.0 }
]
}
```
Notes:
- `timestamp_offset` is added to each word’s `start`/`end` in the response.
- Implementations may perform VAD-based chunking and batching for long-form audio; word timings are adjusted accordingly.
Example curl:
```bash
curl -X POST \
-H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
-H "Content-Type: application/json" \
-d '{
"audio_file_url": "https://example.com/audio.mp3",
"language": "en",
"timestamp_offset": 0
}' \
"$BASE_URL/v1/audio/transcriptions-from-url"
```
### Error handling
- 400 Bad Request
- Parakeet: `language` other than `en`
- Missing required parameters (`file`/`files` for upload; `audio_file_url` for URL endpoint)
- Unsupported file extension
- 401 Unauthorized
- Missing or invalid Bearer token
- 404 Not Found
- `audio_file_url` does not exist
### Implementation details
- GPUs: A10G for small-file/live, L40S for large-file URL transcription (subject to deployment)
- VAD chunking and segment batching; word timings adjusted and overlapping ends constrained
- Pads very short segments (< 0.5s) to avoid model crashes on some backends
### Server configuration (Reflector API)
Set the Reflector server to use the Modal backend and point `TRANSCRIPT_URL` to your chosen deployment:
```
TRANSCRIPT_BACKEND=modal
TRANSCRIPT_URL=https://<account>--reflector-transcriber-parakeet-web.modal.run
TRANSCRIPT_MODAL_API_KEY=<REFLECTOR_GPU_APIKEY>
```
### Conformance tests
Use the pytest-based conformance tests to validate any new implementation (including self-hosted) against this spec:
```
TRANSCRIPT_URL=https://<your-deployment-base> \
TRANSCRIPT_MODAL_API_KEY=your-api-key \
uv run -m pytest -m gpu_modal --no-cov server/tests/test_gpu_modal_transcript.py
```

569
server/gpu/modal_deployments/reflector_transcriber.py
View File

@@ -1,41 +1,78 @@
import os
import tempfile
import sys
import threading
import uuid
from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
from urllib.parse import urlparse
import modal
from pydantic import BaseModel
MODELS_DIR = "/models"
MODEL_NAME = "large-v2"
MODEL_COMPUTE_TYPE: str = "float16"
MODEL_NUM_WORKERS: int = 1
MINUTES = 60 # seconds
SAMPLERATE = 16000
UPLOADS_PATH = "/uploads"
CACHE_PATH = "/models"
SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
VAD_CONFIG = {
"batch_max_duration": 30.0,
"silence_padding": 0.5,
"window_size": 512,
}
volume = modal.Volume.from_name("models", create_if_missing=True)
WhisperUniqFilename = NewType("WhisperUniqFilename", str)
AudioFileExtension = NewType("AudioFileExtension", str)
app = modal.App("reflector-transcriber")
model_cache = modal.Volume.from_name("models", create_if_missing=True)
upload_volume = modal.Volume.from_name("whisper-uploads", create_if_missing=True)
class TimeSegment(NamedTuple):
"""Represents a time segment with start and end times."""
start: float
end: float
class AudioSegment(NamedTuple):
"""Represents an audio segment with timing and audio data."""
start: float
end: float
audio: any
class TranscriptResult(NamedTuple):
"""Represents a transcription result with text and word timings."""
text: str
words: list["WordTiming"]
class WordTiming(TypedDict):
"""Represents a word with its timing information."""
word: str
start: float
end: float
def download_model():
from faster_whisper import download_model
volume.reload()
model_cache.reload()
download_model(MODEL_NAME, cache_dir=MODELS_DIR)
download_model(MODEL_NAME, cache_dir=CACHE_PATH)
volume.commit()
model_cache.commit()
image = (
modal.Image.debian_slim(python_version="3.12")
.pip_install(
"huggingface_hub==0.27.1",
"hf-transfer==0.1.9",
"torch==2.5.1",
"faster-whisper==1.1.1",
)
.env(
{
"HF_HUB_ENABLE_HF_TRANSFER": "1",
@@ -45,19 +82,98 @@ image = (
),
}
)
.run_function(download_model, volumes={MODELS_DIR: volume})
.apt_install("ffmpeg")
.pip_install(
"huggingface_hub==0.27.1",
"hf-transfer==0.1.9",
"torch==2.5.1",
"faster-whisper==1.1.1",
"fastapi==0.115.12",
"requests",
"librosa==0.10.1",
"numpy<2",
"silero-vad==5.1.0",
)
.run_function(download_model, volumes={CACHE_PATH: model_cache})
)
def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
parsed_url = urlparse(url)
url_path = parsed_url.path
for ext in SUPPORTED_FILE_EXTENSIONS:
if url_path.lower().endswith(f".{ext}"):
return AudioFileExtension(ext)
content_type = headers.get("content-type", "").lower()
if "audio/mpeg" in content_type or "audio/mp3" in content_type:
return AudioFileExtension("mp3")
if "audio/wav" in content_type:
return AudioFileExtension("wav")
if "audio/mp4" in content_type:
return AudioFileExtension("mp4")
raise ValueError(
f"Unsupported audio format for URL: {url}. "
f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
)
def download_audio_to_volume(
audio_file_url: str,
) -> tuple[WhisperUniqFilename, AudioFileExtension]:
import requests
from fastapi import HTTPException
response = requests.head(audio_file_url, allow_redirects=True)
if response.status_code == 404:
raise HTTPException(status_code=404, detail="Audio file not found")
response = requests.get(audio_file_url, allow_redirects=True)
response.raise_for_status()
audio_suffix = detect_audio_format(audio_file_url, response.headers)
unique_filename = WhisperUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
file_path = f"{UPLOADS_PATH}/{unique_filename}"
with open(file_path, "wb") as f:
f.write(response.content)
upload_volume.commit()
return unique_filename, audio_suffix
def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
"""Add 0.5s of silence if audio is shorter than the silence_padding window.
Whisper does not require this strictly, but aligning behavior with Parakeet
avoids edge-case crashes on extremely short inputs and makes comparisons easier.
"""
import numpy as np
audio_duration = len(audio_array) / sample_rate
if audio_duration < VAD_CONFIG["silence_padding"]:
silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
silence = np.zeros(silence_samples, dtype=np.float32)
return np.concatenate([audio_array, silence])
return audio_array
@app.cls(
gpu="A10G",
timeout=5 * MINUTES,
scaledown_window=5 * MINUTES,
allow_concurrent_inputs=6,
image=image,
volumes={MODELS_DIR: volume},
volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
)
class Transcriber:
@modal.concurrent(max_inputs=10)
class TranscriberWhisperLive:
"""Live transcriber class for small audio segments (A10G).
Mirrors the Parakeet live class API but uses Faster-Whisper under the hood.
"""
@modal.enter()
def enter(self):
import faster_whisper
@@ -71,23 +187,200 @@ class Transcriber:
device=self.device,
compute_type=MODEL_COMPUTE_TYPE,
num_workers=MODEL_NUM_WORKERS,
download_root=MODELS_DIR,
download_root=CACHE_PATH,
local_files_only=True,
)
print(f"Model is on device: {self.device}")
@modal.method()
def transcribe_segment(
self,
audio_data: str,
audio_suffix: str,
language: str,
filename: str,
language: str = "en",
):
with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
fp.write(audio_data)
"""Transcribe a single uploaded audio file by filename."""
upload_volume.reload()
file_path = f"{UPLOADS_PATH}/{filename}"
if not os.path.exists(file_path):
raise FileNotFoundError(f"File not found: {file_path}")
with self.lock:
with NoStdStreams():
segments, _ = self.model.transcribe(
file_path,
language=language,
beam_size=5,
word_timestamps=True,
vad_filter=True,
vad_parameters={"min_silence_duration_ms": 500},
)
segments = list(segments)
text = "".join(segment.text for segment in segments).strip()
words = [
{
"word": word.word,
"start": round(float(word.start), 2),
"end": round(float(word.end), 2),
}
for segment in segments
for word in segment.words
]
return {"text": text, "words": words}
@modal.method()
def transcribe_batch(
self,
filenames: list[str],
language: str = "en",
):
"""Transcribe multiple uploaded audio files and return per-file results."""
upload_volume.reload()
results = []
for filename in filenames:
file_path = f"{UPLOADS_PATH}/{filename}"
if not os.path.exists(file_path):
raise FileNotFoundError(f"Batch file not found: {file_path}")
with self.lock:
with NoStdStreams():
segments, _ = self.model.transcribe(
file_path,
language=language,
beam_size=5,
word_timestamps=True,
vad_filter=True,
vad_parameters={"min_silence_duration_ms": 500},
)
segments = list(segments)
text = "".join(seg.text for seg in segments).strip()
words = [
{
"word": w.word,
"start": round(float(w.start), 2),
"end": round(float(w.end), 2),
}
for seg in segments
for w in seg.words
]
results.append(
{
"filename": filename,
"text": text,
"words": words,
}
)
return results
@app.cls(
gpu="L40S",
timeout=15 * MINUTES,
image=image,
volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
)
class TranscriberWhisperFile:
"""File transcriber for larger/longer audio, using VAD-driven batching (L40S)."""
@modal.enter()
def enter(self):
import faster_whisper
import torch
from silero_vad import load_silero_vad
self.lock = threading.Lock()
self.use_gpu = torch.cuda.is_available()
self.device = "cuda" if self.use_gpu else "cpu"
self.model = faster_whisper.WhisperModel(
MODEL_NAME,
device=self.device,
compute_type=MODEL_COMPUTE_TYPE,
num_workers=MODEL_NUM_WORKERS,
download_root=CACHE_PATH,
local_files_only=True,
)
self.vad_model = load_silero_vad(onnx=False)
@modal.method()
def transcribe_segment(
self, filename: str, timestamp_offset: float = 0.0, language: str = "en"
):
import librosa
import numpy as np
from silero_vad import VADIterator
def vad_segments(
audio_array,
sample_rate: int = SAMPLERATE,
window_size: int = VAD_CONFIG["window_size"],
) -> Generator[TimeSegment, None, None]:
"""Generate speech segments as TimeSegment using Silero VAD."""
iterator = VADIterator(self.vad_model, sampling_rate=sample_rate)
start = None
for i in range(0, len(audio_array), window_size):
chunk = audio_array[i : i + window_size]
if len(chunk) < window_size:
chunk = np.pad(
chunk, (0, window_size - len(chunk)), mode="constant"
)
speech = iterator(chunk)
if not speech:
continue
if "start" in speech:
start = speech["start"]
continue
if "end" in speech and start is not None:
end = speech["end"]
yield TimeSegment(
start / float(SAMPLERATE), end / float(SAMPLERATE)
)
start = None
iterator.reset_states()
upload_volume.reload()
file_path = f"{UPLOADS_PATH}/{filename}"
if not os.path.exists(file_path):
raise FileNotFoundError(f"File not found: {file_path}")
audio_array, _sr = librosa.load(file_path, sr=SAMPLERATE, mono=True)
# Batch segments up to ~30s windows by merging contiguous VAD segments
merged_batches: list[TimeSegment] = []
batch_start = None
batch_end = None
max_duration = VAD_CONFIG["batch_max_duration"]
for segment in vad_segments(audio_array):
seg_start, seg_end = segment.start, segment.end
if batch_start is None:
batch_start, batch_end = seg_start, seg_end
continue
if seg_end - batch_start <= max_duration:
batch_end = seg_end
else:
merged_batches.append(TimeSegment(batch_start, batch_end))
batch_start, batch_end = seg_start, seg_end
if batch_start is not None and batch_end is not None:
merged_batches.append(TimeSegment(batch_start, batch_end))
all_text = []
all_words = []
for segment in merged_batches:
start_time, end_time = segment.start, segment.end
s_idx = int(start_time * SAMPLERATE)
e_idx = int(end_time * SAMPLERATE)
segment = audio_array[s_idx:e_idx]
segment = pad_audio(segment, SAMPLERATE)
with self.lock:
segments, _ = self.model.transcribe(
fp.name,
segment,
language=language,
beam_size=5,
word_timestamps=True,
@@ -96,66 +389,220 @@ class Transcriber:
)
segments = list(segments)
text = "".join(segment.text for segment in segments)
text = "".join(seg.text for seg in segments).strip()
words = [
{"word": word.word, "start": word.start, "end": word.end}
for segment in segments
for word in segment.words
{
"word": w.word,
"start": round(float(w.start) + start_time + timestamp_offset, 2),
"end": round(float(w.end) + start_time + timestamp_offset, 2),
}
for seg in segments
for w in seg.words
]
if text:
all_text.append(text)
all_words.extend(words)
return {"text": text, "words": words}
return {"text": " ".join(all_text), "words": all_words}
def detect_audio_format(url: str, headers: dict) -> str:
from urllib.parse import urlparse
from fastapi import HTTPException
url_path = urlparse(url).path
for ext in SUPPORTED_FILE_EXTENSIONS:
if url_path.lower().endswith(f".{ext}"):
return ext
content_type = headers.get("content-type", "").lower()
if "audio/mpeg" in content_type or "audio/mp3" in content_type:
return "mp3"
if "audio/wav" in content_type:
return "wav"
if "audio/mp4" in content_type:
return "mp4"
raise HTTPException(
status_code=400,
detail=(
f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
),
)
def download_audio_to_volume(audio_file_url: str) -> tuple[str, str]:
import requests
from fastapi import HTTPException
response = requests.head(audio_file_url, allow_redirects=True)
if response.status_code == 404:
raise HTTPException(status_code=404, detail="Audio file not found")
response = requests.get(audio_file_url, allow_redirects=True)
response.raise_for_status()
audio_suffix = detect_audio_format(audio_file_url, response.headers)
unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
file_path = f"{UPLOADS_PATH}/{unique_filename}"
with open(file_path, "wb") as f:
f.write(response.content)
upload_volume.commit()
return unique_filename, audio_suffix
@app.function(
scaledown_window=60,
timeout=60,
allow_concurrent_inputs=40,
timeout=600,
secrets=[
modal.Secret.from_name("reflector-gpu"),
],
volumes={MODELS_DIR: volume},
volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
image=image,
)
@modal.concurrent(max_inputs=40)
@modal.asgi_app()
def web():
from fastapi import Body, Depends, FastAPI, HTTPException, UploadFile, status
from fastapi import (
Body,
Depends,
FastAPI,
Form,
HTTPException,
UploadFile,
status,
)
from fastapi.security import OAuth2PasswordBearer
from typing_extensions import Annotated
transcriber = Transcriber()
transcriber_live = TranscriberWhisperLive()
transcriber_file = TranscriberWhisperFile()
app = FastAPI()
oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
supported_file_types = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
def apikey_auth(apikey: str = Depends(oauth2_scheme)):
if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid API key",
headers={"WWW-Authenticate": "Bearer"},
)
if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
return
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid API key",
headers={"WWW-Authenticate": "Bearer"},
)
class TranscriptResponse(BaseModel):
result: dict
class TranscriptResponse(dict):
pass
@app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
def transcribe(
file: UploadFile,
model: str = "whisper-1",
language: Annotated[str, Body(...)] = "en",
) -> TranscriptResponse:
audio_data = file.file.read()
audio_suffix = file.filename.split(".")[-1]
assert audio_suffix in supported_file_types
file: UploadFile = None,
files: list[UploadFile] | None = None,
model: str = Form(MODEL_NAME),
language: str = Form("en"),
batch: bool = Form(False),
):
if not file and not files:
raise HTTPException(
status_code=400, detail="Either 'file' or 'files' parameter is required"
)
if batch and not files:
raise HTTPException(
status_code=400, detail="Batch transcription requires 'files'"
)
func = transcriber.transcribe_segment.spawn(
audio_data=audio_data,
audio_suffix=audio_suffix,
language=language,
)
result = func.get()
return result
upload_files = [file] if file else files
uploaded_filenames: list[str] = []
for upload_file in upload_files:
audio_suffix = upload_file.filename.split(".")[-1]
if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
raise HTTPException(
status_code=400,
detail=(
f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
),
)
unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
file_path = f"{UPLOADS_PATH}/{unique_filename}"
with open(file_path, "wb") as f:
content = upload_file.file.read()
f.write(content)
uploaded_filenames.append(unique_filename)
upload_volume.commit()
try:
if batch and len(upload_files) > 1:
func = transcriber_live.transcribe_batch.spawn(
filenames=uploaded_filenames,
language=language,
)
results = func.get()
return {"results": results}
results = []
for filename in uploaded_filenames:
func = transcriber_live.transcribe_segment.spawn(
filename=filename,
language=language,
)
result = func.get()
result["filename"] = filename
results.append(result)
return {"results": results} if len(results) > 1 else results[0]
finally:
for filename in uploaded_filenames:
try:
file_path = f"{UPLOADS_PATH}/{filename}"
os.remove(file_path)
except Exception:
pass
upload_volume.commit()
@app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
def transcribe_from_url(
audio_file_url: str = Body(
..., description="URL of the audio file to transcribe"
),
model: str = Body(MODEL_NAME),
language: str = Body("en"),
timestamp_offset: float = Body(0.0),
):
unique_filename, _audio_suffix = download_audio_to_volume(audio_file_url)
try:
func = transcriber_file.transcribe_segment.spawn(
filename=unique_filename,
timestamp_offset=timestamp_offset,
language=language,
)
result = func.get()
return result
finally:
try:
file_path = f"{UPLOADS_PATH}/{unique_filename}"
os.remove(file_path)
upload_volume.commit()
except Exception:
pass
return app
class NoStdStreams:
def __init__(self):
self.devnull = open(os.devnull, "w")
def __enter__(self):
self._stdout, self._stderr = sys.stdout, sys.stderr
self._stdout.flush()
self._stderr.flush()
sys.stdout, sys.stderr = self.devnull, self.devnull
def __exit__(self, exc_type, exc_value, traceback):
sys.stdout, sys.stderr = self._stdout, self._stderr
self.devnull.close()

3
server/tests/test_gpu_modal_transcript.py
View File

@@ -272,6 +272,9 @@ class TestGPUModalTranscript:
for f in temp_files:
Path(f).unlink(missing_ok=True)
@pytest.mark.skipif(
not "parakeet" in get_model_name(), reason="Parakeet only supports English"
)
def test_transcriptions_error_handling(self):
"""Test error handling for invalid requests."""
url = get_modal_transcript_url()