feat(openai): add support for speech translation

README.md

@@ -33,7 +33,7 @@ uv add scope3ai
 |-------------|-----------------|----|-----|------------------|-----------|
 | Anthropic   | ✅              |    |     |                  |           |
 | Cohere      | ✅              |    |     |                  |           |
-| OpenAI      | ✅              | ✅ | ✅  | ✅               |           |
+| OpenAI      | ✅              | ✅ | ✅  | ✅               | ✅        |
 | Huggingface | ✅              | ✅ | ✅  | ✅               | ✅        |
 | LiteLLM     | ✅              |    |     |                  |           |
 | MistralAi   | ✅              |    |     |                  |           |

scope3ai/tracers/openai/instrument.py

@@ -13,6 +13,10 @@
     openai_async_speech_to_text_wrapper,
     openai_speech_to_text_wrapper,
 )
+from .translation import (
+    openai_translation_wrapper,
+    openai_async_translation_wrapper,
+)
 
 
 class OpenAIInstrumentor:
@@ -48,6 +52,16 @@ def __init__(self) -> None:
                 "name": "AsyncTranscriptions.create",
                 "wrapper": openai_async_speech_to_text_wrapper,
             },
+            {
+                "module": "openai.resources.audio.translations",
+                "name": "Translations.create",
+                "wrapper": openai_translation_wrapper,
+            },
+            {
+                "module": "openai.resources.audio.translations",
+                "name": "AsyncTranslations.create",
+                "wrapper": openai_async_translation_wrapper,
+            },
             {
                 "module": "openai.resources.images",
                 "name": "Images.create_variation",

scope3ai/tracers/openai/speech_to_text.py

@@ -2,23 +2,19 @@
 import time
 from typing import Any, Callable, Optional, Union
 
-import openai
 import tiktoken
-from openai.resources.audio.transcriptions import (
-    AsyncTranscriptions,
-    Transcriptions,
-)
-from openai.resources.audio.transcriptions import (
-    Transcription as _Transcription,
-)
-from openai.resources.audio.transcriptions import (
+from openai.resources.audio.transcriptions import AsyncTranscriptions, Transcriptions
+from openai.types.audio.transcription import Transcription as _Transcription
+from openai.types.audio.transcription_verbose import (
     TranscriptionVerbose as _TranscriptionVerbose,
 )
 
 from scope3ai.api.types import ImpactRow, Model, Scope3AIContext, Task
 from scope3ai.constants import PROVIDERS
 from scope3ai.lib import Scope3AI
 
+from .utils import _get_file_audio_duration
+
 PROVIDER = PROVIDERS.OPENAI.value
 
 logger = logging.getLogger("scope3.tracers.openai.speech_to_text")
@@ -36,23 +32,6 @@ class TranscriptionVerbose(_TranscriptionVerbose):
     scope3ai: Optional[Scope3AIContext] = None
 
 
-def _get_file_audio_duration(
-    file: openai._types.FileTypes,
-) -> Optional[float]:
-    try:
-        from mutagen import File
-
-        if isinstance(file, (list, tuple)):
-            file = file[1]
-
-        audio = File(file)
-        if audio is not None and audio.info is not None:
-            return audio.info.length
-    except Exception as e:
-        logger.exception(f"Failed to get audio duration: {e}")
-    return None
-
-
 def _openai_speech_to_text_wrapper(
     response: Any, request_latency: float, kwargs: dict
 ) -> Union[Transcription, TranscriptionVerbose, str]:

scope3ai/tracers/openai/translation.py

@@ -0,0 +1,91 @@
+import logging
+import time
+from typing import Any, Callable, Optional, Union
+
+import tiktoken
+from openai.resources.audio.translations import AsyncTranslations, Translations
+from openai.types.audio.translation import Translation as _Translation
+from openai.types.audio.translation_verbose import (
+    TranslationVerbose as _TranslationVerbose,
+)
+
+from scope3ai.api.types import ImpactRow, Model, Scope3AIContext, Task
+from scope3ai.constants import PROVIDERS
+from scope3ai.lib import Scope3AI
+
+from .utils import _get_file_audio_duration
+
+PROVIDER = PROVIDERS.OPENAI.value
+
+logger = logging.getLogger(__name__)
+
+
+class AnnotatedStr(str):
+    scope3ai: Optional[Scope3AIContext] = None
+
+
+class Translation(_Translation):
+    scope3ai: Optional[Scope3AIContext] = None
+
+
+class TranslationVerbose(_TranslationVerbose):
+    scope3ai: Optional[Scope3AIContext] = None
+
+
+def _openai_translation_wrapper(
+    response: Any, request_latency: float, kwargs: dict
+) -> Union[Translation, TranslationVerbose, AnnotatedStr]:
+    model = kwargs["model"]
+    encoder = tiktoken.get_encoding("cl100k_base")
+
+    if isinstance(response, (_Translation, _TranslationVerbose)):
+        output_tokens = len(encoder.encode(response.text))
+    elif isinstance(response, str):
+        output_tokens = len(encoder.encode(response))
+    else:
+        output_tokens = None
+
+    options = {}
+    duration = _get_file_audio_duration(kwargs["file"])
+    if duration is not None:
+        options["input_audio_seconds"] = int(duration)
+
+    scope3_row = ImpactRow(
+        model=Model(id=model),
+        provider=PROVIDER,
+        output_tokens=output_tokens,
+        request_duration_ms=request_latency,
+        task=Task.translation,
+        **options,
+    )
+    scope3_ctx = Scope3AI.get_instance().submit_impact(scope3_row)
+
+    if isinstance(response, _Translation):
+        result = Translation.model_construct(**response.model_dump())
+    elif isinstance(response, _TranslationVerbose):
+        result = TranslationVerbose.model_construct(**response.model_dump())
+    elif isinstance(response, str):
+        result = AnnotatedStr(str)
+    else:
+        logger.error(f"Unexpected response type: {type(response)}")
+        return response
+    result.scope3ai = scope3_ctx
+    return result
+
+
+def openai_translation_wrapper(
+    wrapped: Callable, instance: Translations, args: Any, kwargs: Any
+) -> Union[Translation, TranslationVerbose, AnnotatedStr]:
+    timer_start = time.perf_counter()
+    response = wrapped(*args, **kwargs)
+    request_latency = (time.perf_counter() - timer_start) * 1000
+    return _openai_translation_wrapper(response, request_latency, kwargs)
+
+
+async def openai_async_translation_wrapper(
+    wrapped: Callable, instance: AsyncTranslations, args: Any, kwargs: Any
+) -> Union[Translation, TranslationVerbose, AnnotatedStr]:
+    timer_start = time.perf_counter()
+    response = await wrapped(*args, **kwargs)
+    request_latency = (time.perf_counter() - timer_start) * 1000
+    return _openai_translation_wrapper(response, request_latency, kwargs)

scope3ai/tracers/openai/utils.py

@@ -0,0 +1,23 @@
+import logging
+from typing import Optional
+
+import openai
+
+logger = logging.getLogger("scope3.tracers.openai")
+
+
+def _get_file_audio_duration(
+    file: openai._types.FileTypes,
+) -> Optional[float]:
+    try:
+        from mutagen import File
+
+        if isinstance(file, (list, tuple)):
+            file = file[1]
+
+        audio = File(file)
+        if audio is not None and audio.info is not None:
+            return audio.info.length
+    except Exception as e:
+        logger.exception(f"Failed to get audio duration: {e}")
+    return None