---
title: 'SPGISpeech 2.0: Financial Speech Corpus'
url: https://www.emergentmind.com/topics/spgispeech-2-0
type: topic
---

# SPGISpeech 2.0: Financial Speech Corpus

Searching arXiv for the cited papers to ground the article in the current records.
I’ll look up the SPGISpeech and SPGISpeech 2.0 records on arXiv.
SPGISpeech 2.0 is a large-scale, professionally transcribed financial speech corpus designed for speaker-tagged transcription and related multi-talker speech tasks. It extends the original SPGISpeech formulation of fully formatted end-to-end speech recognition by preserving orthographic transcripts while adding call-level and speaker-level metadata, speaker-labeled segments, and per-word speaker information. The corpus contains 3,780 additional hours of English corporate earnings-call audio, more than 150,000 unique audio snippets, and 41,593 unique speakers, with snippets of 50–90 seconds and an average duration of 77.3 seconds. The paper characterizes it as the largest speaker recognition / speaker-tagged transcription dataset available in terms of both hours and number of speakers, and reports that fine-tuning with speaker supervision improves both transcription and speaker attribution performance on realistic earnings-call audio [2508.05554].

## 1. Concept and task formulation

SPGISpeech 2.0 is defined around a task that goes beyond conventional ASR: joint transcription and speaker attribution in financial teleconference audio. Its retained core property from the original SPGISpeech is the pairing of audio snippets with fully formatted text transcriptions suitable for end-to-end ASR. In this setting, the targets are orthographic rather than normalized, preserving capitalization, punctuation, normalization conventions, and disfluency handling rather than deferring them to downstream text-only modules [2508.05554].

The central distinction from standard ASR corpora is that SPGISpeech 2.0 is explicitly constructed for speaker-tagged transcription. The paper positions it as supporting speaker-tagged transcription, speaker diarization, speaker recognition / speaker identification, multi-talker ASR, and integrated ASR + diarization. Every snippet has at least two speakers and at least one speaker change, so the dataset is not merely a transcription corpus with auxiliary speaker labels; the speaker structure is part of the primary learning problem.

This task formulation inherits the orthographic emphasis of the original SPGISpeech, which proposed fully formatted end-to-end speech recognition as an alternative to pipelines that first emit uncased, punctuation-free text and then recover orthography by post-processing. In the earlier corpus, the motivation was that capitalization, punctuation, denormalization of non-standard words, and disfluency treatment can depend on information present in the acoustic signal, including prosodic cues unavailable to text-only normalization components [2104.02014]. SPGISpeech 2.0 preserves that premise while expanding the target problem to include speaker attribution.

## 2. Relation to the original SPGISpeech

The original SPGISpeech corpus consisted of 5,000 hours of professionally transcribed earnings-call audio released for fully formatted end-to-end speech recognition. Its data units were short 5–15 second 16-bit, 16 kHz mono WAV slices paired with fully formatted orthographic transcripts, and its baseline Conformer systems showed that such orthographic prediction was feasible at low error rates, including a reported CER of 1.7 for the ESPnet Conformer on the private test split [2104.02014].

SPGISpeech 2.0 preserves several defining characteristics of that earlier resource. It remains based on English earnings-call audio, uses professional human transcription, follows the same transcription style guide, retains orthographic conventions such as capitalization, punctuation, normalization, and handling of disfluencies, and continues to reflect varied teleconference recording conditions, diverse global L1/L2 accents, and business/finance content rich in named entities [2508.05554].

The major differences are structural. SPGISpeech 2.0 replaces the short ASR-centric slicing regime with 50–90 second snippets, making longer conversational context available for diarization and multi-talker modeling. It adds per-word speaker information, speaker-segment alignments, call metadata, and speaker metadata. It also introduces a much larger speaker inventory, with speaker IDs represented as anonymized integers consistent across calls or as `-1` for unknown speakers. In addition, the paper states that the new release contains both the original human-annotated transcription and an algorithmically adjusted, more literal version intended to better reflect the spoken audio [2508.05554].

The longer snippet design materially changes the acoustic and textual character of the corpus. The paper notes that the new data include more names, numerals, style-guide-driven edits, disfluencies, and small talk, and that the corpus is less “clean” than the original because many segments are drawn from messier Q&A sections. This suggests that SPGISpeech 2.0 is not only an extension in scale and annotation richness, but also a shift toward harder long-context meeting-style conditions.

## 3. Corpus construction, annotation, and transcript variants

SPGISpeech 2.0 is built from professionally transcribed earnings calls. All audio is 16 kHz and single-channel. Audio snippets are extracted from calls and aligned using Gentle forced alignment, py-webrtc voice activity detection, and cross-checking with the NeMo forced alignment pipeline. The stated purpose of this alignment procedure is accurate word-level timing, especially at the start and end of calls [2508.05554].

Speaker annotation is represented at multiple granularities. In transcripts, each speaker change is marked with the pipe character `|`. Speaker labels also appear in SegLST speaker-alignment files and in JSON word-level alignment files. Speaker IDs are anonymized and consistent across calls, while unknown speakers are represented as `-1`. These design choices make the corpus directly usable for models that operate at transcript level, segment level, or word level.

A distinctive feature is the inclusion of two transcript versions. The human-annotated transcription follows the company’s professional style guide and is explicitly described as not always literal: it may omit small talk, filler words, stutters, or corrections. The algorithmically adjusted transcription is intended to be more literal and is created by running an ASR model trained on literal transcriptions, aligning the output to the gold annotations, using numeric-value cross-referencing for denormalization, and reintroducing disfluencies and removed small talk where possible [2508.05554]. This distinction matters because it separates style-guide fidelity from acoustic literalness, which are often conflated in ASR corpora.

The selection criteria further constrain the released data. The paper lists the following inclusion conditions: few non-standard characters or currency symbols; snippets can cover no more than 35% of a given call; snippets must be well aligned for at least five words at both the start and end under both Gentle and NeMo; known speakers were prioritized over unknown speakers; at least 50% of speakers in dev and test appear in train; and mean and median of at least four snippets per speaker [2508.05554]. These conditions imply a deliberate balance between alignment reliability, speaker continuity, and split usability for supervised speaker-aware modeling.

## 4. Splits, scale, and distributional properties

The corpus contains 3,780 hours of audio and more than 150,000 snippets. Snippets are 50–90 seconds long, with an average duration of 77.3 seconds. The speaker inventory comprises 41,593 unique speakers, with an average of 8.1 snippets per speaker when unknown speakers are ignored. Coverage spans all 11 GICS sectors, and the paper reports a global speaker distribution estimated from a sample of 2,000 random speakers [2508.05554].

| Split | Snippets | Calls |
|---|---:|---:|
| Train | 154,971 | 48,000 |
| Dev | 6,447 | 2,000 |
| Test | 7,177 | 2,131 |

A call appears in only one split. Dev and test are the most recent calls chronologically, all training calls occur before dev, and dev occurs before test. The paper states that this avoids chronological leakage [2508.05554]. For financial speech, where firm-specific terminology, recurring participants, and market-period effects can be temporally clustered, this split design is consequential.

The speaker structure within snippets is also reported. Most snippets contain 2 speakers, some have up to 7 speakers, and one snippet has more than six speakers. Speaker talk-time distribution is long-tailed. Region estimates based on corporate domicile for 2,000 random speakers are 49% North America, 24% Asia/Pacific, 20% Europe, 5% Latin America/Caribbean, and 2% Africa/Middle East [2508.05554].

The paper compares entity-like textual prevalence in SPGISpeech 2.0 against the original SPGISpeech using Flair NER on a subset of 2,500 snippets:

| Feature | Original SPGISpeech | SPGISpeech 2.0 |
|---|---:|---:|
| Acronyms | 15% | 23% |
| Pauses | 10% | 66% |
| Organizations | 25% | 52% |
| Persons | 8% | 45% |
| Locations | 8% | 23% |

These prevalence shifts are consistent with the move from short clipped utterances to longer multi-speaker exchanges. A plausible implication is that SPGISpeech 2.0 places greater stress on long-range discourse tracking, entity consistency, and robust handling of speaker turns than the original 5–15 second regime [2104.02014].

## 5. Enabled tasks, benchmark models, and evaluation

The dataset is explicitly intended for speaker-tagged transcription, multi-talker ASR, speaker diarization, speaker recognition / identification, and integrated ASR + diarization. The paper uses two benchmark model families to validate utility: a Canary-based ASR model and a Sortformer-based integrated diarization + ASR model [2508.05554].

Canary is described as an encoder-decoder ASR model using a FastConformer encoder and Transformer decoder. Sortformer extends Canary with a parallel encoder for diarization, a sort loss to resolve speaker-permutation ambiguity, and end-to-end training for diarization + ASR. The experimental framing is therefore not limited to post hoc diarization layered on top of ASR; it also includes jointly optimized architectures.

Evaluation uses standard WER and speaker-aware cpWER. WER is defined as

$$
\mathrm{WER} = \frac{S + D + I}{N},
$$

where \(S\), \(D\), and \(I\) denote substitutions, deletions, and insertions, and \(N\) is the number of reference words. For speaker-tagged transcription, the paper reports concatenated minimum permutation word error rate:

$$
\mathrm{cpWER} = \min_{\pi \in \Pi} \mathrm{WER}\big(\text{concat}(\hat{y}_{\pi}), y\big),
$$

that is, WER after choosing the optimal permutation of predicted speaker streams to reference speaker streams. Results are reported both with punctuation and capitalization included (“With PnC”) and with punctuation and capitalization removed (“Without-PnC”) [2508.05554].

The reported strongest system is Canary-170M + Sortformer-123M fine-tuned on SPGISpeech 2.0 with speaker supervision:

| Setting | cpWER | WER |
|---|---:|---:|
| With PnC | 15.88% | 7.25% |
| Without-PnC | 13.39% | 4.62% |

For context, the paper states that baseline systems include Canary-170M, Canary-1B, Whisper-turbo, and Whisper-large-v3, and that these systems generally perform worse without SPGISpeech 2.0-specific fine-tuning. It reports for baseline Canary-170M a cpWER of 20.53% and WER of 7.96%, and states that Whisper systems perform notably worse on this task, especially in speaker-tagged settings [2508.05554].

The paper’s main empirical conclusion is that speaker supervision matters: fine-tuning on SPGISpeech 2.0 improves both ASR and speaker-tagging quality, and the joint ASR + diarization model benefits from the speaker annotations. In that sense, the corpus functions not only as training data but also as a benchmark for end-to-end speaker-aware financial speech recognition.

## 6. Limitations, release conditions, and research role

The paper is explicit that SPGISpeech 2.0 is not a balanced general-purpose speech dataset. Its primary limitation is finance-domain specificity: topics are mostly earnings-call related. It also notes demographic skew because the speaker population is biased toward people who appear on financial calls. Word-level alignments are available, but they are algorithmic rather than manually produced. Overlapping speech is relatively limited because earnings-call conventions reduce overlap, so the corpus may be less challenging than more naturally interruptive conversational settings. The release also excludes some low-quality snippets, especially those where prior ASR, diarization, or speaker systems performed in the bottom decile [2508.05554].

These limitations are important for interpretation. Strong results on SPGISpeech 2.0 do not directly establish robustness on casual conversation, heavily overlapping meetings, or non-financial domains. At the same time, the corpus targets a class of professionally consequential speech settings that are underrepresented in public data: long-form teleconference audio with speaker changes, specialized terminology, and orthographic transcription requirements.

The dataset is released free for non-commercial use. The paper states that the public link is redacted in the camera-ready text and “will not be public before the end of the review period” [2508.05554]. The original SPGISpeech was likewise released free for non-commercial use and made available as a contribution to the STT research community [2104.02014].

Within that lineage, SPGISpeech 2.0 occupies a specific position. The original corpus established that fully formatted end-to-end transcription of financial speech was viable at scale. SPGISpeech 2.0 extends that agenda from orthographic ASR to speaker-attributed transcription by combining longer context, richer metadata, and explicit speaker supervision. Its most direct research significance lies in narrowing the gap between large transcript-only ASR corpora and speaker datasets that lack realistic orthographic transcriptions, scale, or domain richness.

Source: https://www.emergentmind.com/topics/spgispeech-2-0