Skip to content

vllm.entrypoints.speech_to_text.transcription.protocol

Classes:

TranscriptionDiarizedSegment

Bases: OpenAIBaseModel

A speaker-attributed transcription segment.

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
402
403
404
405
406
407
408
409
410
class TranscriptionDiarizedSegment(OpenAIBaseModel):
    """A speaker-attributed transcription segment."""

    type: Literal["transcript.text.segment"] = "transcript.text.segment"
    id: str
    start: float
    end: float
    text: str
    speaker: str

TranscriptionRequest

Bases: OpenAIBaseModel

Attributes:

  • file (UploadFile) –

    The audio file object (not file name) to transcribe, in one of these

  • frequency_penalty (float | None) –

    The frequency penalty to use for sampling.

  • hotwords (str | None) –

    hotwords refers to a list of important words or phrases that the model

  • include_stop_str_in_output (bool) –

    Whether to include the stop strings in output text.

  • language (str | None) –

    The language of the input audio.

  • length_penalty (float) –

    Length penalty to be used for beam search.

  • max_completion_tokens (int | None) –

    The maximum number of tokens to generate.

  • min_p (float | None) –

    Filters out tokens with a probability lower than min_p, ensuring a

  • model (str | None) –

    ID of the model to use.

  • n (int) –

    The number of beams to be used in beam search.

  • presence_penalty (float | None) –

    The presence penalty to use for sampling.

  • prompt (str) –

    An optional text to guide the model's style or continue a previous audio

  • repetition_penalty (float | None) –

    The repetition penalty to use for sampling.

  • response_format (TranscriptionResponseFormat) –

    The format of the output, in one of these options: json, text, srt,

  • seed (int | None) –

    The seed to use for sampling.

  • stream (bool | None) –

    When set, it will enable output to be streamed in a similar fashion

  • temperature (float) –

    The sampling temperature, between 0 and 1.

  • timestamp_granularities (list[Literal['word', 'segment']]) –

    The timestamp granularities to populate for this transcription.

  • to_language (str | None) –

    The language of the output audio we transcribe to.

  • top_k (int | None) –

    Limits sampling to the k most probable tokens at each step.

  • top_p (float | None) –

    Enables nucleus (top-p) sampling, where tokens are selected from the

  • use_beam_search (bool) –

    Whether or not beam search should be used.

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
class TranscriptionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/audio/createTranscription

    file: UploadFile
    """
    The audio file object (not file name) to transcribe, in one of these
    formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm.
    """

    model: str | None = None
    """ID of the model to use.
    """

    language: str | None = None
    """The language of the input audio.

    Supplying the input language in
    [ISO-639-1](https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes) format
    will improve accuracy and latency.
    """

    hotwords: str | None = None
    """
    hotwords refers to a list of important words or phrases that the model
    should pay extra attention to during transcription.
    """

    prompt: str = Field(default="")
    """An optional text to guide the model's style or continue a previous audio
    segment.

    The [prompt](https://platform.openai.com/docs/guides/speech-to-text#prompting)
    should match the audio language.
    """

    response_format: TranscriptionResponseFormat = Field(default="json")
    """
    The format of the output, in one of these options: `json`, `text`, `srt`,
    `verbose_json`, or `vtt`.
    """

    ## TODO (varun) : Support if set to 0, certain thresholds are met !!

    timestamp_granularities: list[Literal["word", "segment"]] = Field(
        alias="timestamp_granularities[]", default=[]
    )
    """The timestamp granularities to populate for this transcription.

    `response_format` must be set `verbose_json` to use timestamp granularities.
    Either or both of these options are supported: `word`, or `segment`. Note:
    There is no additional latency for segment timestamps, but generating word
    timestamps incurs additional latency.
    """

    stream: bool | None = False
    """When set, it will enable output to be streamed in a similar fashion
    as the Chat Completion endpoint.
    """
    # --8<-- [start:transcription-extra-params]
    # Flattened stream option to simplify form data.
    stream_include_usage: bool | None = False
    stream_continuous_usage_stats: bool | None = False

    vllm_xargs: dict[str, str | int | float | list[str | int | float]] | None = Field(
        default=None,
        description=(
            "Additional request parameters with (list of) string or "
            "numeric values, used by custom extensions."
        ),
    )
    # --8<-- [end:transcription-extra-params]

    to_language: str | None = None
    """The language of the output audio we transcribe to.

    Please note that this is not currently used by supported models at this
    time, but it is a placeholder for future use, matching translation api.
    """

    # --8<-- [start:transcription-sampling-params]
    use_beam_search: bool = False
    """Whether or not beam search should be used."""

    n: int = 1
    """The number of beams to be used in beam search."""

    length_penalty: float = 1.0
    """Length penalty to be used for beam search."""

    include_stop_str_in_output: bool = False
    """Whether to include the stop strings in output text."""

    temperature: float = Field(default=0.0)
    """The sampling temperature, between 0 and 1.

    Higher values like 0.8 will make the output more random, while lower values
    like 0.2 will make it more focused / deterministic. If set to 0, the model
    will use [log probability](https://en.wikipedia.org/wiki/Log_probability)
    to automatically increase the temperature until certain thresholds are hit.
    """

    top_p: float | None = None
    """Enables nucleus (top-p) sampling, where tokens are selected from the
    smallest possible set whose cumulative probability exceeds `p`.
    """

    top_k: int | None = None
    """Limits sampling to the `k` most probable tokens at each step."""

    min_p: float | None = None
    """Filters out tokens with a probability lower than `min_p`, ensuring a
    minimum likelihood threshold during sampling.
    """

    seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
    """The seed to use for sampling."""

    frequency_penalty: float | None = 0.0
    """The frequency penalty to use for sampling."""

    repetition_penalty: float | None = None
    """The repetition penalty to use for sampling."""

    presence_penalty: float | None = 0.0
    """The presence penalty to use for sampling."""

    max_completion_tokens: int | None = None
    """The maximum number of tokens to generate."""
    # --8<-- [end:transcription-sampling-params]

    # Default sampling parameters for transcription requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
        "top_k": 0,
        "min_p": 0.0,
    }

    def build_stt_params(
        self,
        audio: "np.ndarray",
        stt_config: "SpeechToTextConfig",
        model_config: "ModelConfig",
        task_type: str,
    ) -> SpeechToTextParams:
        return SpeechToTextParams(
            audio=audio,
            stt_config=stt_config,
            model_config=model_config,
            language=self.language,
            task_type=task_type,
            request_prompt=self.prompt,
            to_language=self.to_language,
            hotwords=self.hotwords,
        )

    def to_beam_search_params(
        self,
        default_max_tokens: int,
        default_sampling_params: dict | None = None,
    ) -> BeamSearchParams:
        if default_sampling_params is None:
            default_sampling_params = {}

        max_tokens = default_max_tokens
        n = self.n if self.n is not None else 1

        # NOTE: Temp 0 is a different fallback than completions
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get("temperature", 0)

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            temperature=temperature,
            length_penalty=self.length_penalty,
            include_stop_str_in_output=self.include_stop_str_in_output,
        )

    def to_sampling_params(
        self, default_max_tokens: int, default_sampling_params: dict | None = None
    ) -> SamplingParams:
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}

        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]
            )
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]
            )
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
            )

        if (repetition_penalty := self.repetition_penalty) is None:
            repetition_penalty = default_sampling_params.get(
                "repetition_penalty",
                self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"],
            )

        return SamplingParams.from_optional(
            temperature=temperature,
            max_tokens=max_tokens,
            seed=self.seed,
            top_p=top_p,
            top_k=top_k,
            min_p=min_p,
            frequency_penalty=self.frequency_penalty,
            repetition_penalty=repetition_penalty,
            presence_penalty=self.presence_penalty,
            output_kind=RequestOutputKind.DELTA
            if self.stream
            else RequestOutputKind.FINAL_ONLY,
            extra_args=self.vllm_xargs,
            skip_clone=True,  # Created fresh per request, safe to skip clone
        )

    @model_validator(mode="before")
    @classmethod
    def validate_transcription_request(cls, data):
        if not isinstance(data, dict):
            return data
        if isinstance(data.get("file"), str):
            raise HTTPException(
                status_code=HTTPStatus.UNPROCESSABLE_ENTITY,
                detail="Expected 'file' to be a file-like object, not 'str'.",
            )

        stream_opts = ["stream_include_usage", "stream_continuous_usage_stats"]
        stream = data.get("stream", False)
        if any(bool(data.get(so, False)) for so in stream_opts) and not stream:
            # Find which specific stream option was set
            invalid_param = next(
                (so for so in stream_opts if data.get(so, False)),
                "stream_include_usage",
            )
            raise VLLMValidationError(
                "Stream options can only be defined when `stream=True`.",
                parameter=invalid_param,
            )

        # Parse vllm_xargs from JSON string (form data sends it as a string)
        xargs = data.get("vllm_xargs")
        if isinstance(xargs, str):
            try:
                data["vllm_xargs"] = json.loads(xargs)
            except json.JSONDecodeError as e:
                raise VLLMValidationError(
                    f"Failed to parse vllm_xargs. Must be valid JSON: {e}",
                    parameter="vllm_xargs",
                ) from e

        return data

file instance-attribute

The audio file object (not file name) to transcribe, in one of these formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm.

frequency_penalty = 0.0 class-attribute instance-attribute

The frequency penalty to use for sampling.

hotwords = None class-attribute instance-attribute

hotwords refers to a list of important words or phrases that the model should pay extra attention to during transcription.

include_stop_str_in_output = False class-attribute instance-attribute

Whether to include the stop strings in output text.

language = None class-attribute instance-attribute

The language of the input audio.

Supplying the input language in ISO-639-1 format will improve accuracy and latency.

length_penalty = 1.0 class-attribute instance-attribute

Length penalty to be used for beam search.

max_completion_tokens = None class-attribute instance-attribute

The maximum number of tokens to generate.

min_p = None class-attribute instance-attribute

Filters out tokens with a probability lower than min_p, ensuring a minimum likelihood threshold during sampling.

model = None class-attribute instance-attribute

ID of the model to use.

n = 1 class-attribute instance-attribute

The number of beams to be used in beam search.

presence_penalty = 0.0 class-attribute instance-attribute

The presence penalty to use for sampling.

prompt = Field(default='') class-attribute instance-attribute

An optional text to guide the model's style or continue a previous audio segment.

The prompt should match the audio language.

repetition_penalty = None class-attribute instance-attribute

The repetition penalty to use for sampling.

response_format = Field(default='json') class-attribute instance-attribute

The format of the output, in one of these options: json, text, srt, verbose_json, or vtt.

seed = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max) class-attribute instance-attribute

The seed to use for sampling.

stream = False class-attribute instance-attribute

When set, it will enable output to be streamed in a similar fashion as the Chat Completion endpoint.

temperature = Field(default=0.0) class-attribute instance-attribute

The sampling temperature, between 0 and 1.

Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused / deterministic. If set to 0, the model will use log probability to automatically increase the temperature until certain thresholds are hit.

timestamp_granularities = Field(alias='timestamp_granularities[]', default=[]) class-attribute instance-attribute

The timestamp granularities to populate for this transcription.

response_format must be set verbose_json to use timestamp granularities. Either or both of these options are supported: word, or segment. Note: There is no additional latency for segment timestamps, but generating word timestamps incurs additional latency.

to_language = None class-attribute instance-attribute

The language of the output audio we transcribe to.

Please note that this is not currently used by supported models at this time, but it is a placeholder for future use, matching translation api.

top_k = None class-attribute instance-attribute

Limits sampling to the k most probable tokens at each step.

top_p = None class-attribute instance-attribute

Enables nucleus (top-p) sampling, where tokens are selected from the smallest possible set whose cumulative probability exceeds p.

Whether or not beam search should be used.

TranscriptionResponse

Bases: OpenAIBaseModel

Attributes:

  • text (str) –

    The transcribed text.

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
326
327
328
329
class TranscriptionResponse(OpenAIBaseModel):
    text: str
    """The transcribed text."""
    usage: TranscriptionUsageAudio

text instance-attribute

The transcribed text.

TranscriptionResponseDiarized

Bases: OpenAIBaseModel

OpenAI-compatible diarized transcription response.

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
413
414
415
416
417
418
419
420
class TranscriptionResponseDiarized(OpenAIBaseModel):
    """OpenAI-compatible diarized transcription response."""

    task: Literal["transcribe"] = "transcribe"
    duration: float
    text: str
    segments: list[TranscriptionDiarizedSegment]
    usage: TranscriptionUsageAudio

TranscriptionResponseVerbose

Bases: OpenAIBaseModel

Attributes:

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
class TranscriptionResponseVerbose(OpenAIBaseModel):
    duration: float
    """The duration of the input audio."""

    language: str
    """The language of the input audio."""

    text: str
    """The transcribed text."""

    segments: list[TranscriptionSegment] | None = None
    """Segments of the transcribed text and their corresponding details."""

    words: list[TranscriptionWord] | None = None
    """Extracted words and their corresponding timestamps."""

duration instance-attribute

The duration of the input audio.

language instance-attribute

The language of the input audio.

segments = None class-attribute instance-attribute

Segments of the transcribed text and their corresponding details.

text instance-attribute

The transcribed text.

words = None class-attribute instance-attribute

Extracted words and their corresponding timestamps.

TranscriptionSegment

Bases: OpenAIBaseModel

Attributes:

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
class TranscriptionSegment(OpenAIBaseModel):
    id: int
    """Unique identifier of the segment."""

    avg_logprob: float
    """Average logprob of the segment.

    If the value is lower than -1, consider the logprobs failed.
    """

    compression_ratio: float
    """Compression ratio of the segment.

    If the value is greater than 2.4, consider the compression failed.
    """

    end: float
    """End time of the segment in seconds."""

    no_speech_prob: float | None = None
    """Probability of no speech in the segment.

    If the value is higher than 1.0 and the `avg_logprob` is below -1, consider
    this segment silent.
    """

    seek: int
    """Seek offset of the segment."""

    start: float
    """Start time of the segment in seconds."""

    temperature: float
    """Temperature parameter used for generating the segment."""

    text: str
    """Text content of the segment."""

    tokens: list[int]
    """Array of token IDs for the text content."""

avg_logprob instance-attribute

Average logprob of the segment.

If the value is lower than -1, consider the logprobs failed.

compression_ratio instance-attribute

Compression ratio of the segment.

If the value is greater than 2.4, consider the compression failed.

end instance-attribute

End time of the segment in seconds.

id instance-attribute

Unique identifier of the segment.

no_speech_prob = None class-attribute instance-attribute

Probability of no speech in the segment.

If the value is higher than 1.0 and the avg_logprob is below -1, consider this segment silent.

seek instance-attribute

Seek offset of the segment.

start instance-attribute

Start time of the segment in seconds.

temperature instance-attribute

Temperature parameter used for generating the segment.

text instance-attribute

Text content of the segment.

tokens instance-attribute

Array of token IDs for the text content.

TranscriptionWord

Bases: OpenAIBaseModel

Attributes:

  • end (float) –

    End time of the word in seconds.

  • start (float) –

    Start time of the word in seconds.

  • word (str) –

    The text content of the word.

Source code in vllm/entrypoints/speech_to_text/transcription/protocol.py 
332
333
334
335
336
337
338
339
340
class TranscriptionWord(OpenAIBaseModel):
    end: float
    """End time of the word in seconds."""

    start: float
    """Start time of the word in seconds."""

    word: str
    """The text content of the word."""

end instance-attribute

End time of the word in seconds.

start instance-attribute

Start time of the word in seconds.

word instance-attribute

The text content of the word.