Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

Harden AssemblyAI connector against cross-connector settings: ignore shared TRANSCRIPTION_BASE_URL/MODEL, validate models against AssemblyAI's set, and use the plural speech_models field

Murtaza Nasircc4c2350fbc3
diff --git a/config/env.transcription.example b/config/env.transcription.example
index 26821a03..80653f1b 100644
--- a/config/env.transcription.example
+++ b/config/env.transcription.example
@@ -176,8 +176,17 @@ TRANSCRIPTION_MODEL=gpt-4o-transcribe-diarize
# TRANSCRIPTION_CONNECTOR=assemblyai
# TRANSCRIPTION_API_KEY=your_assemblyai_api_key
#
-# Optional: pin a speech model (omit to use your account default).
-# TRANSCRIPTION_MODEL=universal
+# The endpoint is fixed at https://api.assemblyai.com, so the shared
+# TRANSCRIPTION_BASE_URL is intentionally ignored for this connector (a leftover
+# value from another connector won't misroute uploads). Set ASSEMBLYAI_BASE_URL
+# only if you front the API with a proxy.
+# ASSEMBLYAI_BASE_URL=
+#
+# Optional: pin a speech model (omit to use your account default). Uses a
+# dedicated var, NOT the shared TRANSCRIPTION_MODEL (a leftover model name from
+# another connector would be rejected by AssemblyAI). Comma-separated values
+# become an ordered fallback list (AssemblyAI's speech_models).
+# ASSEMBLYAI_SPEECH_MODEL=universal-3-pro,universal-2
#
# Diarization, language, and hotwords come from each user's settings (or tag /
# folder defaults). Hotwords are sent as AssemblyAI word_boost; min/max speaker
diff --git a/src/services/transcription/connectors/assemblyai.py b/src/services/transcription/connectors/assemblyai.py
index d36cdae9..a62bad2d 100644
--- a/src/services/transcription/connectors/assemblyai.py
+++ b/src/services/transcription/connectors/assemblyai.py
@@ -10,7 +10,8 @@ chunking is needed.
Config (from env via the registry):
- api_key: TRANSCRIPTION_API_KEY (required; sent as the `authorization` header, no Bearer)
- base_url: TRANSCRIPTION_BASE_URL (default https://api.assemblyai.com)
-- model: TRANSCRIPTION_MODEL (optional; passed as `speech_model`, omitted when blank)
+- model: ASSEMBLYAI_SPEECH_MODEL (optional; one or comma-separated AssemblyAI models
+ -> `speech_models`; foreign/blank values fall back to the account default)
"""
import io
@@ -45,6 +46,14 @@ class AssemblyAITranscriptionConnector(BaseTranscriptionConnector):
}
PROVIDER_NAME = "assemblyai"
+ # AssemblyAI's own speech models. The transcription-model that reaches a
+ # connector can be resolved from shared settings (env, admin default, tag /
+ # folder defaults) that were meant for a DIFFERENT connector (e.g. WhisperX
+ # 'large-v3' or OpenAI 'gpt-4o-transcribe-diarize'). Sending one of those to
+ # AssemblyAI is a hard error, so we only forward models AssemblyAI knows and
+ # otherwise fall back to the account default.
+ KNOWN_MODELS = frozenset({'universal-3-pro', 'universal-2', 'universal', 'nano', 'best', 'slam-1'})
+
# AssemblyAI ingests long files in a single async job (10h / up to 2.2GB on
# upload), so the app must not chunk for this connector.
SPECIFICATIONS = ConnectorSpecifications(
@@ -126,10 +135,19 @@ class AssemblyAITranscriptionConnector(BaseTranscriptionConnector):
payload["word_boost"] = words
payload["boost_param"] = self.boost_param
- # Optional model override (account default when blank).
+ # Optional model selection (account default when blank). AssemblyAI
+ # deprecated the singular `speech_model` in favour of `speech_models`
+ # (an ordered fallback array). Only forward models AssemblyAI recognises;
+ # a foreign model name resolved from shared settings is dropped so we
+ # fall back to the account default rather than erroring.
model = self._effective_model(request)
if model:
- payload["speech_model"] = model
+ candidates = [m.strip() for m in model.split(',') if m.strip()]
+ valid = [m for m in candidates if m.lower() in self.KNOWN_MODELS]
+ if valid:
+ payload["speech_models"] = valid
+ elif candidates:
+ logger.info("AssemblyAI: ignoring non-AssemblyAI model(s) %s; using account default", candidates)
return payload
@@ -162,8 +180,8 @@ class AssemblyAITranscriptionConnector(BaseTranscriptionConnector):
audio_url = self._upload(client, audio_bytes)
payload = self._build_payload(request, audio_url)
logger.info(
- "AssemblyAI: submitting transcript (diarize=%s, lang=%s, model=%s)",
- request.diarize, payload.get('language_code', 'auto'), payload.get('speech_model', 'default'))
+ "AssemblyAI: submitting transcript (diarize=%s, lang=%s, models=%s)",
+ request.diarize, payload.get('language_code', 'auto'), payload.get('speech_models', 'default'))
resp = client.post("/v2/transcript", json=payload,
headers={"content-type": "application/json"})
if resp.status_code not in (200, 201):
diff --git a/src/services/transcription/registry.py b/src/services/transcription/registry.py
index 292ace07..05e3d4d2 100644
--- a/src/services/transcription/registry.py
+++ b/src/services/transcription/registry.py
@@ -292,15 +292,20 @@ class ConnectorRegistry:
}
elif connector_name == 'assemblyai':
- base_url = os.environ.get('TRANSCRIPTION_BASE_URL', '')
- if base_url:
- base_url = base_url.split('#')[0].strip()
+ # AssemblyAI's endpoint is fixed (api.assemblyai.com), so we do NOT
+ # read the shared TRANSCRIPTION_BASE_URL here: a stale value left
+ # over from another connector would otherwise send uploads to the
+ # wrong host. Use the dedicated ASSEMBLYAI_BASE_URL only for proxies.
+ base_url = os.environ.get('ASSEMBLYAI_BASE_URL', '').split('#')[0].strip()
return {
'api_key': os.environ.get('TRANSCRIPTION_API_KEY', ''),
'base_url': base_url or 'https://api.assemblyai.com',
- # Optional speech_model; account default when blank.
- 'model': os.environ.get('TRANSCRIPTION_MODEL', ''),
+ # Dedicated model var (NOT the shared TRANSCRIPTION_MODEL, which
+ # is often a leftover model name from another connector that
+ # AssemblyAI would reject). Account default when blank; a
+ # comma-separated list becomes an ordered speech_models fallback.
+ 'model': os.environ.get('ASSEMBLYAI_SPEECH_MODEL', ''),
}
else: # openai_whisper (default)
diff --git a/tests/test_cov_assemblyai_connector.py b/tests/test_cov_assemblyai_connector.py
index 8a4ea794..b91bc729 100644
--- a/tests/test_cov_assemblyai_connector.py
+++ b/tests/test_cov_assemblyai_connector.py
@@ -127,9 +127,32 @@ def test_payload_no_diarize_no_speaker_fields():
assert p['language_code'] == 'de'
-def test_payload_model_passthrough():
- p = _conn(model='universal')._build_payload(_req(diarize=False, language='en'), 'u')
- assert p['speech_model'] == 'universal'
+def test_payload_model_passthrough_uses_plural_array():
+ # Single model -> one-element speech_models array (plural field).
+ p = _conn(model='universal-3-pro')._build_payload(_req(diarize=False, language='en'), 'u')
+ assert p['speech_models'] == ['universal-3-pro']
+ assert 'speech_model' not in p
+
+
+def test_payload_model_comma_list_becomes_fallback_array():
+ p = _conn(model='universal-3-pro, universal-2')._build_payload(_req(diarize=False, language='en'), 'u')
+ assert p['speech_models'] == ['universal-3-pro', 'universal-2']
+
+
+def test_payload_no_model_omits_speech_models():
+ p = _conn()._build_payload(_req(diarize=False, language='en'), 'u')
+ assert 'speech_models' not in p and 'speech_model' not in p
+
+
+def test_payload_foreign_model_dropped():
+ # A model name resolved from shared settings for another connector
+ # (WhisperX 'large-v3', OpenAI 'gpt-4o-...') must NOT be sent to AssemblyAI.
+ for bad in ('large-v3', 'gpt-4o-transcribe-diarize'):
+ p = _conn(model=bad)._build_payload(_req(diarize=False, language='en'), 'u')
+ assert 'speech_models' not in p, f"{bad} should have been dropped"
+ # A per-request override that's foreign is dropped too.
+ p = _conn(model='universal-2')._build_payload(_req(diarize=False, language='en', model='large-v3'), 'u')
+ assert 'speech_models' not in p
# --- end-to-end flow (mocked) ---------------------------------------------