Sign Up

gRPC API Reference

Basic Transcription

These requests are used to transcribe audio synchronously. The Transcribe request is suitable for transcription of short audio, while the TranscribeStream request is suitable for transcription of possibly long audio as well as real-time audio.

rpc Transcribe

rpc Transcribe(TranscribeRequest) returns (TranscribeResponse) {}

message TranscribeRequest {
  string api_key = 1;
  TranscriptionConfig config = 4;
  bytes audio = 3;
}

message TranscribeResponse {
  Result result = 1;
  repeated Result channel_results = 2;
}

The Transcribe request transcribes the provided audio and returns the complete transcription at once.

Transcription configuration is specified using the config field in the request; refer to TranscriptionConfig.

Audio data is provided in the audio field. By default, the audio is assumed to use a container format, which is inferred. For raw PCM formats, the specific format, sample rate and number of channels must be specified. For more information about formats and related configuration parameters, refer to Audio Format.

The maximum size of the audio field is 5 MB, while the maximum audio duration is 60 seconds. Exceeding these limits will result in an error and no transcription.

The result of the transcription is returned in the result field of the response as a Result message. However, if separate recognition per channel is enabled, results for consecutive audio channels are returned in the channel_results field, and the result field is not present.

rpc TranscribeStream

rpc TranscribeStream(stream TranscribeStreamRequest) returns (stream TranscribeStreamResponse) {}

message TranscribeStreamRequest {
  string api_key = 1;
  TranscriptionConfig config = 4;
  bytes audio = 3;
}

message TranscribeStreamResponse {
  Result result = 1;
}

The TranscribeStream request transcribes audio in streaming mode, either optimized for throughput or low-latency.

Transcription configuration is specified using the config field in the first request; refer to TranscriptionConfig.

The client sends a sequence of TranscribeStreamRequest requests to the service. The api_key and config are specified in the first request and must not be present in later requests. The audio is provided in chunks using the audio field, which may be empty or non-empty in any of the requests.

TranscribeStream supports the same audio formats as Transcribe; refer to Audio Format. If a container format is used, data from chunks in consecutive requests is effectively concatenated (the precise locations where the client splits the audio into chunks is not important). However, if one of the supported raw PCM formats is used, then each audio chunk must contain a whole number of frames (a frame is a sequence of samples for each channel).

The maximum size of the audio field in a single request is 5 MB, while the maximum total audio duration is 5 hours. If the audio field is too large, an error will be returned immediately. If the maximum total audio duration is exceeded, audio up to the maximum duration will be processed and then an error will be returned.

The service returns a sequence of TranscribeStreamResponse messages, with transcription results available in the result field. The result field may not be present, and it is important for the client to check for its presence before interpreting it (e.g. response.has_result() in C++, response.HasField("result") in Python). This is important for correct handling of non-final words (see below).

The client should not make any assumptions about the correspondence of requests to responses or the presence of result. Even if it appears that the service generates responses in a specific manner, there are no guarantees that any such nonspecified behavior would be maintained.

If config.include_nonfinal is false, TranscribeStream returns only final words and does not offer any latency guarantees. This mode is optimized for throughput and is essentially a streaming version of Transcribe, enabling transcription of longer audio. The complete sequence of words is obtained by joining the words from all results.

If config.include_nonfinal is true, TranscribeStream returns both final and non-final words while minimizing the recognition latency. This mode is suitable for transcription of real-time audio where transcribed words need to be received as soon possible after the associated audio has been sent to the service. The current transcription from the start of the audio can be constructed by joining (in order) final words from results before the last result and all words from the last result. For more information, refer to Final vs Non-final Words.

If config.include_nonfinal is true, the following also applies:

  • Audio should not be sent at a rate faster than real-time. If it is, the service may throttle processing or return an error. There are margins such that this should not occur for real-time streams under normal circumstances.
  • Minimum latency is achieved when using the PCM format pcm_s16le with sample rate 16 kHz and one audio channel. Alternatively, any supported PCM format can be used with a negligible effect on latency. Using a container format in this mode is not recommended due to the latency introduced by audio decoding.

If separate recognition per channel is enabled, different audio channels are transcribed independently (as if each channel was transcribed with its own TranscribeStream request). The result.channel field indicates which channel a result is for; consulting this field is essential to correctly interpret the results. There are no guarantees about the relative order or timing of results for different channels.

Transcription Configuration

message TranscriptionConfig

message TranscriptionConfig {
  // Input options
  string audio_format = 1;
  int32 sample_rate_hertz = 2;
  int32 num_audio_channels = 3;

  // Output options
  bool include_nonfinal = 4;
  bool enable_separate_recognition_per_channel = 16;
  bool enable_endpoint_detection = 18;

  // Speech adaptation
  SpeechContext speech_context = 5;

  // Content moderation
  bool enable_profanity_filter = 6;
  repeated string content_moderation_phrases = 7;

  // Speaker diarization
  bool enable_streaming_speaker_diarization = 8;
  bool enable_global_speaker_diarization = 9;
  int32 min_num_speakers = 10;
  int32 max_num_speakers = 11;

  // Speaker identification
  bool enable_speaker_identification = 12;
  repeated string cand_speaker_names = 13;

  // Model options
  string model = 14;
  bool enable_dictation = 15;

  // Asynchronous transcription
  string transcribe_async_mode = 17;
}

The TranscriptionConfig message is used with all transcription requests and specifies various configuration parameters.

The audio_format, sample_rate_hertz and num_audio_channels fields specify information about the input audio. Refer to Audio Format.

include_nonfinal specifies whether to enable low-latency recognition and include non-final words in results. It is only valid for TranscribeStream. Also refer to Final vs Non-final Words.

enable_separate_recognition_per_channel specifies whether to perform separate speech recognition for each audio channel. When used with Transcribe, a separate result for each channel is explicitly returned in the response. When used with TranscribeStream, results for different channels are multiplexed in the response stream, and result.channel must be used to associate the result to a specific channel. The same applies when retrieving the result of an asynchronous transcription using GetTranscribeAsyncResult.

enable_endpoint_detection enables endpoint detection for interactive voice applications, When the end of the utterance is detected, an <end> word is returned, and all words up to and including the <end> word are returned as final. This feature is designed to be used together with the IVR Domain Model, but can also be used with other models.

speech_context is used to specify the custom vocabulary; refer to Custom Vocabulary for general information. There are two methods of specifying a speech context. First, speech context entries can be included directly in speech_context.entries. Second, a speech context stored in Soniox Cloud (in the context of the user account) can be referenced using speech_context.name; the API for managing stored speech contexts is described in Speech Context Management. It is not allowed to specify both entries and name.

enable_profanity_filter enables the profanity filter to mask profane words/phrases, while content_moderation_phrases specifies custom words/phrases that should be masked. These features can be used on their own or together, and both result in specific words being masked. Masking is performed such that all characters in a word except the first are replaced by asterisks (for example, f***); the original words can still be obtained using the orig_text field in Word. Words in content_moderation_phrases may contain only the following characters: lower-case English letters, - and '.

For fields related to speaker diarization and identification, refer to Speaker AI.

The model field specifies the model to use; refer to Model Type and Medical Domain. Valid models are precision, enhanced, base and precision_medical. If not specified, the default is precision.

For TranscribeAsync, transcribe_async_mode specifies the asynchronous transcription mode (instant_file or sameday_file). If not specified, the default is instant_file. Refer to Soniox Pricing.

Transcription Results

message Result

message Result {
  repeated Word words = 1;
  int32 final_proc_time_ms = 2;
  int32 total_proc_time_ms = 3;
  repeated ResultSpeaker speakers = 6;
  int32 channel = 7;
}

message ResultSpeaker {
  int32 speaker = 1;
  string name = 2;
}

The Result message represents a speech recognition result, containing transcribed words and other data.

The words field contains a sequence of Word messages representing transcribed words.

The final_proc_time_ms and total_proc_time_ms fields determine the duration of processed audio in milliseconds, resulting in final and all words respectively. In a Transcribe request, both values are equal. In a TranscribeStream request, these are consistent with the timestamps of words, such that final words are in the interval from 0 to final_proc_time_ms and non-final words are in the interval from final_proc_time_ms to total_proc_time_ms. These values never decrease in subsequent results for the same transcription.

If using Speaker Identification, the speakers field contains the latest associations between speaker numbers and candidate speakers (for all words from the start of the audio, not just words in this result).

When separate recognition per channel is enabled, the channel field indicates the audio channel that the result is associated with. Audio channels are numbered starting with 0.

message Word

message Word {
  string text = 1;
  int32 start_ms = 2;
  int32 duration_ms = 3;
  bool is_final = 4;
  int32 speaker = 5;
  string orig_text = 8;
  double confidence = 9;
}

The Word message represents an individual recognized word, which is given in the text field.

Punctuation symbols are represented as individual. When converting a transcription result to text for human consumption, it is important to recognize punctuation symbols and not add a space between the previous word and the punctuation symbol. For this purpose, the following text values should be treated as punctuation symbols: ,, ;, ., ?, !.

start_ms and duration_ms represent the time interval of the word in the audio. These are consistent with the order of words, such that start_ms of the next word is always greater than or equal to start_ms + duration_ms of the current word.

is_final specifies if the word is final. This disinction is relevant only when using TranscribeStream with include_nonfinal equal to true; in other cases is_final is always true. Refer to Final vs Non-final Words.

The orig_text field indicates the original word in if the word in text is masked. Refer to Profanity Filter and Custom Content Moderation.

Profanity Filter.

If using Speaker Diarization, the speaker field indicates the speaker number. Valid speaker numbers are greater than or equal to 1.

The confidence field specifies the confidence, in the range between 0 and 1. The value of 0 means that confidence information is not available for the word (this may be the case for punctuation words).

Asynchronous Transcription

These requests allow a client to upload files to be transcribed asynchronously and to retrieve the transcription results later. This feature supports a variety of media file formats including video files.

A file is uploaded for transcription using TranscribeAsync. The status of the transcription can be queried using GetTranscribeAsyncStatus. The result of the transcription is retrieved using GetTranscribeAsyncResult.

A file that has been uploaded (and not yet deleted) is in one of the following states:

  • QUEUED: The file is queued to be transcribed.
  • TRANSCRIBING: The file is being transcribed.
  • COMPLETED: The file has been transcribed successfully, the result is available.
  • FAILED: Transcription has failed, the result is not and will not be available.

A file that is not in the TRANSCRIBING state can be deleted using DeleteTranscribeAsyncFile. It is the responsibility of the user to delete files, they are not deleted automatically.

There is a limit on the numbers of files that have been uploaded but not yet deleted, which is 100. It the limit is reached, further TranscribeAsync requests will be rejected with gRPC status code RESOURCE_EXHAUSTED and details message starting with <too_many_files>. It is the responsibility of the user to prevent or handle these errors.

rpc TranscribeAsync

rpc TranscribeAsync(stream TranscribeAsyncRequest) returns (TranscribeAsyncResponse) {}

message TranscribeAsyncRequest {
  string api_key = 1;
  string reference_name = 3;
  TranscriptionConfig config = 5;
  bytes audio = 4;
}

message TranscribeAsyncResponse {
  string file_id = 1;
}

The TranscribeAsync request is used to upload a file for asyncronous transcription. The client sends a sequence of TranscribeAsyncRequest messages, where the first message specifies the api_key, reference_name and config and may contain an audio chunk, and any further messages contain only an audio chunk. The audio chunks are concatenated to form the complete audio file. The maximum size of an audio chunk is 5 MB. The maximum total size is 500 MB. The maximum total duration of audio is 5 hours.

The asynchronous transcription mode (instant_file or sameday_file) can be specified in config.transcribe_async_mode. If not specified, the default is instant_file. Refer to Soniox Pricing.

Audio is extracted from the file as a part of TranscribeAsync. For larger files, it may take a few seconds to decode the audio after all the audio data has been received by the service. If there is an error during decoding, the TranscribeAsync request will fail with gRPC status code UNKNOWN and details message starting with <invalid_media_file>.

The reference_name specified in the first request is intended to enable the user to identify the file after it has been uploaded. It can be any string not longer than 256 characters, including the empty string, and duplicates are allowed. The service does not use this field, it is only for reference.

If TranscribeAsync succeeds, the automatically assigned file_id is returned. The current state of the file can be queried by calling GetTranscribeAsyncStatus with the file_id.

The file will initially be in the QUEUED state and will transition to the TRANSCRIBING state when the transcription starts. The time that this takes depends on the current service load. When transcription has completed, the file will transition to the COMPLETED state and the results can be retrieved using GetTranscribeAsyncResult. If transcription fails, the file will instead transition to the FAILED state. The file will then remain in the COMPLETED or FAILED state until it is deleted using DeleteTranscribeAsyncFile

rpc GetTranscribeAsyncStatus

rpc GetTranscribeAsyncStatus(GetTranscribeAsyncStatusRequest) returns (GetTranscribeAsyncStatusResponse) {}

message GetTranscribeAsyncStatusRequest {
  string api_key = 1;
  string file_id = 2;
}

message GetTranscribeAsyncStatusResponse {
  repeated TranscribeAsyncFileStatus files = 1;
}

message TranscribeAsyncFileStatus {
  string file_id = 1;
  string reference_name = 2;
  // One of: QUEUED, TRANSCRIBING, COMPLETED, FAILED
  string status = 3;
  // UTC timestamp
  google.protobuf.Timestamp created_time = 4;
  string error_message = 5;
  string transcribe_async_mode = 6;
}

The GetTranscribeAsyncStatus request returns the state and other information for a specific file or for all existing files.

If file_id in the request is non-empty, information for a file with that ID is returned. In this case, if there is no file with that ID, the request fails with gRPC status code NOT_FOUND and details message starting with <file_id_not_found>. If file_id is empty, information about all existing files is returned.

The files field in the response is a sequence of TranscribeAsyncFileStatus messages. If file_id in the request was non-empty, there will be exactly one element, otherwise there will be one element for each existing file ordered by increasing created_time.

The following information is returned for each file in the response file_id, reference_name, status (state), created_time (UTC timestamp when TranscribeAsync has completed), error_message (error information if the status is FAILED) and transcribe_async_mode (instant_file or sameday_file).

rpc GetTranscribeAsyncResult

rpc GetTranscribeAsyncResult(GetTranscribeAsyncResultRequest) returns (stream GetTranscribeAsyncResultResponse) {}

message GetTranscribeAsyncResultRequest {
  string api_key = 1;
  string file_id = 2;
}

message GetTranscribeAsyncResultResponse {
  bool separate_recognition_per_channel = 2;
  Result result = 1;
}

The GetTranscribeAsyncResult request retrieves the the transcription results for a file in the COMPLETED state.

The file for which to retrieve results is specified by file_id. If there is no file with that ID, the request fails with gRPC status code NOT_FOUND and details message starting with <file_id_not_found>.

If the file is still in the QUEUED or TRANSCRIBING state, the request fails with gRPC status code FAILED_PRECONDITION and details message starting with <file_not_transcribed_yet>. If the file is in the FAILED state, it fails with gRPC status code FAILED_PRECONDITION and details message starting with <file_transcription_failed>.

The transcription results are returned as a sequence of Result messages embedded in a sequence of GetTranscribeAsyncResultResponse responses, similar to TranscribeStream. The user can assemble the complete result by concatenating the words from all responses, which are guaranteed to be final, and taking the fields final_proc_time_ms, total_proc_time_ms and speakers from the last response.

If separate recognition per channel was enabled, as is indicated by the field separate_recognition_per_channel in each response (having the same value in all responses), the assembly of results described above must be done on a per-channel basis according to result.channel in each response.

rpc DeleteTranscribeAsyncFile

rpc DeleteTranscribeAsyncFile(DeleteTranscribeAsyncFileRequest) returns (DeleteTranscribeAsyncFileResponse) {}

message DeleteTranscribeAsyncFileRequest {
  string api_key = 1;
  string file_id = 2;
}

message DeleteTranscribeAsyncFileResponse {
}

The DeleteTranscribeAsyncFile request deletes a specific file.

The file to delete is specified by file_id. If there is no file with that ID, the request fails with gRPC status code NOT_FOUND and details message starting with <file_id_not_found>.

A file can be deleted as long as it is not in the TRANSCRIBING state. If it is, the request fails with gRPC status code FAILED_PRECONDITION and details message starting with <file_being_transcribed>.

Transcription of Meetings

The TranscribeMeeting request is provided for the purpose of real-time low-latency transcription of a meeting with a separate audio stream for each participant. A requirement for using this is that the application performs voice activity detection and sends only segments of audio with voice activity detected.

The term stream is used as synonymous with meeting participant. The term segment means a contiguous audio recording sent to the service in the context of a specific stream. Each segment is itself sent to the service in a number of requests, to enable low-latency operation.

rpc TranscribeMeeting

rpc TranscribeMeeting(stream TranscribeMeetingRequest) returns (stream TranscribeMeetingResponse) {}

message TranscribeMeetingRequest {
  string api_key = 1;
  TranscriptionConfig config = 10;
  int32 seq_num = 3;
  int32 stream_id = 4;
  bool start_of_segment = 5;
  bytes audio = 6;
  bool end_of_segment = 7;
}

message TranscribeMeetingResponse {
  int32 seq_num = 1;
  int32 stream_id = 2;
  bool start_of_segment = 3;
  bool end_of_segment = 4;
  Result result = 5;
  string error = 6;
}

The client sends a sequence of TranscribeMeetingRequest requests to the service. The api_key and config are provided in the first request and must not be included in later requests. The configuration specified in config applies to all streams. Since TranscribeMeeting is intended only for the real-time low-latency use case, it is required that config.include_nonfinal is set to true.

The seq_num is an opaque value which is returned in each response. More information about responses is given below.

The stream_id determines which stream the fields start_of_segment, audio and end_of_segment apply to. A stream_id of 0 means no stream, and in that case these fields must have default values. It is recommended to send at least one request every 10 seconds to prevent the session from timing out; using stream_id 0 enables doing so when no audio needs to be sent to the service.

Assuming that stream_id is not 0, the audio field contains new audio data for the stream, if any. The start_of_segment and end_of_segment flags indicate whether an audio segment starts before audio, or ends after audio, respectively. These flags must be consistent within a stream. Specifically: start_of_segment must be true in the first request for the stream, and start_of_segment must also be true if end_of_segment was true in the previous request for the same stream. If this is not the case, some of the audio will not be processed.

Note that, effectively, an audio segment is defined as the concatenation of audio data starting from a request where start_of_segment is true up to first request (the same or a later one) where end_of_segment is true, considering only requests for the same stream.

IMPORTANT: Each audio segment is decoded into audio samples independently. If using a container format, each audio segment must be encoded independently of previous segments in the same stream.

IMPORTANT: Active streams, that is streams where end_of_segment was false in the last request for that stream, occupy resources on the service. Make sure to terminate an active stream when it is no longer relevant by sending a request with end_of_segment equal to true (for example, when the meeting participant disconnects).

Transcription results are returned as a stream of TranscribeMeetingResponse responses. For each request, the service will send exactly one response, which will have the same seq_num, stream_id, start_of_segment and end_of_segment. Within the same stream, the order of responses will match the order of requests, but this is not generally true within different streams. The seq_num field can be used to reliably match responses to requests.

IMPORTANT: In the future, the behavior may change such that there might not be one response for each request, but one response could represent a number of consecutive requests for the same stream belonging to the same segment. In such a response, start_of_segment would be that of the first of these requests, while end_of_segment and seq_num would be that of the last of these requests.

Errors specific to a stream generally do not result in the entire TranscribeStream failing, but are reported using the error field in the response. If the error field is non-empty, an error has occured, and the value of the field is the error message. It is important that the client application checks for and reports these errors.

The actual transcription results are given in the result field in the same manner as for TranscribeStream, but they must be interpreted in the context of the specific stream. Note that the client must check whether result is present before interpreting it (refer to TranscribeStream). For the special stream ID 0, result will never be present.

The transcript is always finalized at the end of each segment. Specifically, in a response where end_of_segment is true, result will be present and will not contain any non-final words.

Speech Context Management

These requests are used to manage the user's speech contexts stored in the Soniox Cloud. Storing a speech context enables using it in a Transcribe or TranscribeStream request as an alternative to directly specifying it. For general information about speech contexts, refer to Custom Vocabulary.

Stored speech contexts exist in the context of a user account, where they are uniquely identified by a user-specified name. The user account is inferred from the API key used in the request.

rpc CreateSpeechContext

rpc CreateSpeechContext(CreateSpeechContextRequest) returns (CreateSpeechContextResponse) {}

message CreateSpeechContextRequest {
  string api_key = 1;
  SpeechContext speech_context = 2;
}

message CreateSpeechContextResponse {
}

The CreateSpeechContext request creates a stored speech context.

The name of the speech context to create is specified as a part of the speech context in speech_context.name, which must be non-empty.

If a speech context with that name already exists, an error with status code ALREADY_EXISTS is returned. If the speech_context does not satisfy the SpeechContext requirements, an error with status code INVALID_ARGUMENT and a message describing the problem is returned.

rpc UpdateSpeechContext

rpc UpdateSpeechContext(UpdateSpeechContextRequest) returns (UpdateSpeechContextResponse) {}

message UpdateSpeechContextRequest {
  string api_key = 1;
  SpeechContext speech_context = 2;
}

message UpdateSpeechContextResponse {
}

The UpdateSpeechContext request updates an existing stored speech context.

The name of the speech context to update is specified as a part of the speech context in speech_context.name.

If there is no speech context with that name, an error with status code NOT_FOUND is returned. If the speech_context does not satisfy the SpeechContext requirements, an error with status code INVALID_ARGUMENT and a message describing the problem is returned.

rpc DeleteSpeechContext

rpc DeleteSpeechContext(DeleteSpeechContextRequest) returns (DeleteSpeechContextResponse) {}

message DeleteSpeechContextRequest {
  string api_key = 1;
  string name = 2;
}

message DeleteSpeechContextResponse {
}

The DeleteSpeechContext deletes a stored speech context.

The name of the speech context to delete is specified in name. If there is no speech context with that name, an error with status code NOT_FOUND is returned.

rpc ListSpeechContextNames

rpc ListSpeechContextNames(ListSpeechContextNamesRequest) returns (ListSpeechContextNamesResponse) {}

message ListSpeechContextNamesRequest {
  string api_key = 1;
}

message ListSpeechContextNamesResponse {
  repeated string names = 1;
}

The ListSpeechContextNames request returns the names of all stored speech contexts.

The names are returned in the names field in no specific order.

rpc GetSpeechContext

rpc GetSpeechContext(GetSpeechContextRequest) returns (GetSpeechContextResponse) {}

message GetSpeechContextRequest {
  string api_key = 1;
  string name = 2;
}

message GetSpeechContextResponse {
  SpeechContext speech_context = 1;
}

The GetSpeechContext request retrieves a stored speech context.

The name of the speech context to retrieve is sepecified in the name field. If there is no speech context with that name, an error with status code NOT_FOUND is returned.

message SpeechContext

message SpeechContext {
  repeated SpeechContextEntry entries = 1;
  string name = 2;
}

The SpeechContext message represents a speech context.

The entries field contains the entries defining the speech context, represented by SpeechContextEntry messages. The name field represents the name of the speech context.

The presence of entries and name depends on the context:

  • When a speech context is sent in CreateSpeechContext or UpdateSpeechContext, or returned in GetSpeechContext, both are required or guaranteed to be non-empty respectively.
  • In a Transcribe or TranscribeStream request, either both or exactly one must be non-empty. If both are empty, no speech context is used; if entries is non-empty, these entries are used; if name is non-empty, a stored speech context with that name is used.

Requirements:

  • The size of the name must not exceed 50 characters.
  • For each entry: SpeechContextEntry requirements.
  • The number of phrases in the entire speech context must not exceed 100.
  • There must be no duplicate phrases in the entire speech context, after removing leading, trailing and repeated spaces.

message SpeechContextEntry

message SpeechContextEntry {
  repeated string phrases = 1;
  double boost = 2;
}

The SpeechContextEntry message represents an entry in a speech context, defined by a list of phrases and a single boost value that applies to these phrases. Words in each phrase are separated by spaces.

Requirements:

  • There must be at least phrase.
  • The size of a phrase must not exceed 100 characters.
  • The number of words in a phrase must be between 1 and 5
  • The size of a word must not exceed 25 characters.
  • A phrase may contain only the following characters: a-z (lower-case only), ' (apostrophe), - (hyphen/minus), (space).
  • The boost value must be between -30 and 30 inclusive.

Speaker AI

Speaker Diarization

Speaker diarization distinguishes speakers based on their voice. Please refer to the Speaker AI guide for general information.

Speaker diarization is available for the Transcribe, TranscribeStream and TranscribeAsync. It is enabled by setting config.enable_global_speaker_diarization or config.enable_streaming_speaker_diarization to true, to use global or streaming speaker diarization mode respectively. When speaker diarization is enabled, a speaker number is included with each returned word (speaker field in Word).

When global speaker diarization is used with TranscribeStream, specific restrictions and considerations apply:

  • config.include_nonfinal must be false. Therefore, real-time recognition is not possible.
  • Transcription results will be returned only after the end of the request stream. It may take some time before these are returned, depending on the audio duration.
  • The total audio duration is limited to no more than 2 hours.

Stereaming speaker diarization does not have the restrictions above, but generally has lower accuracy, since it is optimized for low-latency real-time transcription.

When using speaker diarization, the minimum and maximum number of speakers can be specified by setting config.min_num_speakers and config.max_num_speakers respectively. By default (if these values are 0), the service assumes that there are between 1 and 10 speakers. The maximum permitted value of max_num_speakers is 20. Note that if the actual number of speakers in the audio is outside of the specified (or default) range, the accuracy of speaker diarization may be low.

Speaker Identification

Speaker identification works together with speaker diarization to associate numbered speakers with named candidate speakers based on voice samples provided in advance by the user.

A set of gRPC API calls are available for speaker management:

  • AddSpeaker: Add a new speaker.
  • GetSpeaker: Return information about a specific speaker.
  • RemoveSpeaker: Remove a specific speaker.
  • ListSpeakers: Return a list of registered speakers.
  • AddSpeakerAudio: Add a new audio for a specific speaker.
  • GetSpeakerAudio: Retrieve a specific audio of a specific speaker.
  • RemoveSpeakerAudio: Remove a specific audio of a specific speaker.

Each speaker is identified by a speaker name, and each of the speaker's audios is identified by an audio name. Speaker names are unique in the context of the Soniox user account, while audio names are unique in the context of the speaker they belong to.

A simple command-line application manage_speakers is provided as a frontend to the speaker management API. This application can be used to add speakers and audios for testing purposes, and it is also a good reference for using these API calls directly.

In order to use speaker identification with Transcribe or TranscribeStream, the following must be done:

  • Speaker Diarization must be enabled (either global or streaming mode).
  • config.enable_speaker_identification must be set to true.
  • Names of candidate speakers must be provided in config.cand_speaker_names.

Each of the candidate speakers specified must be an existing speaker as added using AddSpeaker (or manage_speakers --add_speaker). If this is not the case, and error will be returned. However, if some of these speakers do not have any audios, no error will be returned, but it will not be possible to identify those speakers.

Results of speaker identification are provided in the speakers field in the Result structure. This is a list of associations between speaker numbers and candidate speakers. This list will not include entries for speaker numbers that were not associated with a candidate speaker.

cookie Change your cookie preferences