Interpreted Call Audio for Speech Translation Models

This guide sits under our speech and audio data buyer's guide and is written for teams building real-time interpretation for support, healthcare, banking or public-service lines. If you already know what you need, you can describe the interpreted conversations you want sourced.

Why interpreted calls fill a gap that public speech translation corpora leave

Public speech translation corpora are mostly read or broadcast speech with offline text translations, so they underrepresent spontaneous, two-party, telephony-band service dialogue. CoVoST 2, for example, covers 21 languages into English and English into 15 languages, but it is built on crowdsourced read sentences from Common Voice [1][4]. Large multitask models such as SeamlessM4T and its streaming successor show what broad coverage can achieve across speech-to-speech, speech-to-text and text-to-speech directions [2][3], yet domain performance on a pharmacy refill line or an insurance claim call still depends on in-domain data.

Interpreted calls differ from monolingual multilingual ASR data, which has one language per recording (see multilingual ASR training data sourcing). They also differ from code-switching audio, where one speaker mixes languages. In an interpreted call, each language is tied to a role, and the interpreter produces a human translation in near real time, which is what a cascaded or end-to-end speech translation system must imitate.

What a usable interpreted call record contains

A usable record separates every turn by speaker role and language and timestamps it, so that each source utterance can be paired with its interpreted rendition. Language diarization (which language is spoken in which segment) is a recognized requirement for multilingual conversational audio [10], and in interpreted calls it should be paired with role diarization: caller, interpreter, agent and sometimes a supervisor or IVR prompt.

The most valuable recordings are multi-channel. Three-way calls from a phone interpreting service often place the interpreter on a separate leg, which lets you separate overlap without relying on a diarization model; the trade-offs are covered in dual-channel call recordings for ASR and diarization. Much of this audio is 8 kHz narrowband G.711 or compressed VoIP, so read telephony vs wideband audio before assuming a wideband model will transfer.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "call_id": "c_000184",
  "domain": "pharmacy_benefits_support",
  "language_pair": ["es-MX", "en-US"],
  "channels": {"ch0": "caller", "ch1": "agent", "ch2": "interpreter"},
  "audio": {"codec": "pcm_s16le", "sample_rate_hz": 8000, "source_codec": "g711_ulaw"},
  "segments": [
    {"seg": 12, "role": "caller", "lang": "es-MX", "start": 41.20, "end": 46.85,
     "text": "[NAME_1] dice que la receta ya no la cubren", "pair_id": "p7"},
    {"seg": 13, "role": "interpreter", "lang": "en-US", "start": 47.10, "end": 50.30,
     "text": "She says her prescription is no longer covered", "pair_id": "p7",
     "rendition_mode": "consecutive", "person_shift": "third_person"},
    {"seg": 14, "role": "agent", "lang": "en-US", "start": 50.90, "end": 55.40,
     "text": "Can I get the member ID to check the formulary", "pair_id": "p8"}
  ],
  "redaction": {"method": "transcript_tag_plus_audio_tone", "qa_sample_checked": true},
  "reference_translation": {"p7": "My prescription is no longer covered"}
}

Note the person_shift field: interpreters on service calls often switch to third person ("she says"), which a first-person translation model should not learn as the target.

Interpreter renditions are not translation references by default

Interpreter output is a valid target for some tasks and a poor reference for others, so decide per use case before you buy. Research comparing simultaneous interpretation with offline translation finds that interpreters compress, drop some content words and reorder to deliver early [5], and simultaneous MT models trained on offline translations lost up to 13.83 BLEU when scored against interpretation data [6]. The Europarl Simultaneous Interpreting Corpus was used to compare translating from the source speaker with translating from the interpreter on both quality and latency [7], and the NAIST English-Japanese corpus found that experienced interpreters handled the latency-quality trade-off better than less experienced ones [8].

For service calls, expect the same patterns plus clarification turns ("Interpreter speaking, may I ask the caller to repeat?"), summarized long answers and third-person framing. Practical consequences:

• Training a streaming model to behave like an interpreter: renditions are the target, ideally filtered by interpreter experience and QA score.
• Measuring translation accuracy: commission separate offline reference translations of the source turns, and keep renditions as a second reference or a human baseline.
• Evaluating simultaneous quality: use metrics designed for interpretation rather than BLEU alone; automatic SI quality metrics are an active research area [9].
• Clarification and meta-turns: tag and exclude them from parallel pairs; keep them for dialogue-management training.

Specification checklist for an interpreted call request

A precise request names language pairs, interpreting mode, domain, channel layout and the alignment depth you need, because these drive whether a supplier's archive is usable at all. Use the template below as the core of a request or a statement of work, and see the general contact-center audio requirements template for fields shared with monolingual calls.

Illustrative example: invented to show structure; it does not describe an available dataset.

Rights in interpreted calls involve three parties

Interpreted calls have at least three participants and often three organizations: the caller, the contact center or provider that owns the line, and the interpreting vendor whose staff or contractor speaks on the call. Before licensing, confirm who controls the recording, whether the interpreting contract allows the recording to be reused for model training, and what notice each party received. The checks in call-recording consent for AI training apply, with the added question of whether the interpreter's voice and performance are covered.

Interpreter renditions are also human translations. If you plan to combine them with text assets such as licensed translation memories, make sure the license defines the transcripts and translations as included records, not only the audio.

Medical interpreting calls carry PHI on both language legs

Medical interpreting recordings contain protected health information in every language spoken, so de-identification must cover both source and target transcripts and both audio streams. HIPAA allows two methods: Safe Harbor, which removes 18 listed identifiers, and Expert Determination, in which a qualified expert finds the re-identification risk very small [11][12]. A limited data set under a data use agreement is a separate route defined in 45 CFR 164.514(e) [12].

Common failure modes in interpreted calls: a name redacted in the English transcript but left in the Spanish one; a date of birth spelled out by the interpreter after the caller's version was masked; or audio bleeps aligned to one channel only. Ask for the redaction method in writing, a per-language QA sample, and evidence that the alignment still holds after masking. For deeper guidance, see redacting spoken PII from call recordings and, for adjacent clinical audio, doctor-patient conversation audio.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Evaluating a sample before you commit

A short, representative sample tells you whether the archive matches your spec far more reliably than a data card. Run these checks on a few hours per language pair:

1. Channel and role check: confirm each channel maps to one role and that interpreter turns are not mixed into the agent channel.
2. Alignment audit: sample 50 pairs per language direction and verify that each rendition actually corresponds to its source turn.
3. Reference gap: have a bilingual reviewer produce offline translations for a subset and measure how far renditions diverge (omissions, additions, third-person shifts).
4. Redaction parity: check that every masked entity in one language is masked in the other, in both text and audio.
5. Baseline run: score an off-the-shelf multilingual model [2] on the sample to see where in-domain data moves the needle; reuse the split as a held-out set, following voice agent evaluation set practice.

How SourceX sources interpreted call audio

SourceX sources operational datasets, including support and sales histories, from US companies and manages the commercial process, including licensing and ongoing purchases. Datasets are sourced on request rather than held in stock, so a request for interpreted calls does not guarantee a match. You describe the data, not the businesses; SourceX looks for US businesses that hold it, and every release is approved by the supplying company.

The process runs Find, Assess (the data and its licensing permissions), Agree (pricing and allowed uses in a license), Transact and Manage, and nothing is contracted until a supplier agrees. Every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Personal details such as names, emails, phone and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect; health records require HIPAA de-identification by Safe Harbor or Expert Determination. Related pages: call center audio datasets and training data for multilingual enterprise models.

Source interpreted conversations for your speech translation model

If your speech translation model needs real interpreted service calls rather than read speech, describe the language pairs, domain, channel layout and reference needs in your request. SourceX serves AI teams wherever they are based and looks for US businesses that hold the data you describe. Start a buyer request.