Voice of Monetary Policy: Gorodnichenko, Pham & Talavera (2023)

What this is. The paper’s core results, the VoiceTone construction and the deep-learning emotion model, and the local-projections specification: enough to know what it found and how, without reading all 37 pages. To replicate or extend it, read the full source at the original or the preprint.

TL;DR

Gorodnichenko, Pham, and Talavera (2023) build a deep learning model that classifies emotions from audio features of Fed chair answers during FOMC press conference Q&A sessions (April 2011 to June 2019) into positive, negative, or neutral. They aggregate these into a VoiceTone measure, then run local projections following Jorda (2005) of daily returns on 14 financial outcomes against VoiceTone and TextSentiment, controlling for Swanson (2021) policy shocks (FFR, FG, AP) and the Wu and Xia (2016) shadow rate. A one-unit increase in VoiceTone raises S&P 500 returns by roughly 100 basis points over five days, reduces VIX, lowers inflation expectations, and appreciates the dollar against the euro. The bond market shows no significant response. Text sentiment (BERT-based) does not have a statistically significant effect on stock prices in this 36-meeting sample. The results are robust to Fed chair fixed effects, alternative text sentiment measures (RoBERTa, FinBERT, search-and-count, human classification per Cieslak, Morse, and Vissing-Jorgensen (2019) style controls), and a high-frequency intraday analysis showing a +1 bp immediate impact on impact. A concurrent paper by Curti and Kazinnik (2021) confirms nonverbal effects through facial expressions.

Core results

Estimates from nonparametric accelerated bootstrap with 90% bias-corrected confidence intervals. Locators point into the source PDF.

#	Result	Locator	Magnitude
R1	Positive voice tone raises S&P 500 (SPY) returns: effect builds over days, significant at 10%	Figure 3 Panel B, p. 562	+~100 bps at h=5 days for +1 unit VoiceTone; +~75 bps per +1 SD (abstract, p. 550)
R2	Positive voice tone reduces stock market volatility (VIX): economically significant negative effect, builds over 15 days	Figure 4 Panel B, p. 563-564	Reduction roughly equal to increase from a 1-SD FG shock (paper’s comparison, p. 563)
R3	Bond market (GOVT ETF) does not respond significantly to voice tone: statistically insignificant at all horizons	Figure 7 Panel B, p. 567	Point estimates fluctuate near zero; 90% CI contains zero throughout h=0-15 days
R4	Positive voice tone raises GOVT-TIP spread (signals lower inflation expectations): near zero on impact, builds and peaks at h~10 days	Figure 10, p. 570	Gradual increase in spread, statistically significant at longer horizons; GLD (gold) price shows similar pattern (Figure 11, p. 571)
R5	Positive voice tone leads to dollar appreciation vs. euro (EUR/USD coefficient positive): builds over 5-10 days	Figure 13, p. 573	~5-7 bps (from figure scale), marginally significant; JPY/USD shows no significant response
R6	Text sentiment (BERT) does not significantly move SPY: positive point estimates but inside CI throughout	Figure 3 Panel C, p. 562	Positive point estimates across all horizons, statistically insignificant in this sample
R7	Intraday (flow spec): positive voice tone raises SPY by +1 bp on impact, answer by answer	Figure 15 top panels, p. 576-577	+1 bp at h=0 minutes, statistically significant; cumulative spec shows gradual build during Q&A session

Overall (paper’s conclusion). Voice tone conveys economically significant information that moves financial markets across multiple asset classes, beyond the content of spoken words and beyond the Fed’s explicit policy actions. The effect builds gradually over days, consistent with information frictions and slow incorporation of nonverbal signals into prices via continued media coverage (Figure 16, p. 578-580).

Theory / model

The paper has no formal structural model. The empirical strategy tests whether voice tone carries market-moving information orthogonal to text and policy actions, under two interpretations that the data cannot distinguish.

Interpretation 1 (forward-guidance signaling, pp. 576-577). A positive voice tone signals that the Fed is unlikely to tighten policy in the near future. Under the zero lower bound, this attenuates rate risk, reduces perceived future volatility (VIX falls), lowers discount uncertainty (equity prices rise), and, since monetary tightening is a lower-probability event, reduces the expected future path of inflation (GOVT-TIP spread rises). Responses to voice tone are qualitatively similar to those for the FG shock from Swanson (2021), consistent with this interpretation.

Interpretation 2 (Fed information effect, Romer and Romer 2000, p. 577). The Fed chair’s positive tone signals superior private information about a stronger economic outlook. A more upbeat assessment reduces uncertainty, may lower the rate-risk component of yields, and raises equity valuations. The paper tests this by checking whether positive-tone press conferences are followed by more dovish Fed tweets on days h=1 to 15; Figure 16 (p. 578-579) confirms this pattern, consistent with information revelation.

Identification. The design controls for observable channels: VoiceTone enters after partialling out TextSentiment (BERT-based, built from statements, remarks, and Q&A responses), three identified policy shocks from Swanson (2021) (FFR, FG, AP), the Wu and Xia (2016) shadow rate, and an indicator for meetings without press conferences. Voice tone and text sentiment are only weakly correlated ($\rho = 0.37$ (Pearson; Spearman = 0.30) for statement text sentiment vs. voice tone; $\rho = 0.29$ for Q&A text sentiment vs. voice tone; pp. 557-558), providing independent variation. Hansen and McMahon (2016) document that FOMC textual content carries significant information; controlling for it isolates the vocal dimension. The design is selection-on-observables: no instrument for voice tone is used, and the identifying assumption is that residual VoiceTone variation is uncorrelated with omitted drivers of financial outcomes after the controls listed above.

Method

Step 1: Emotion detection and VoiceTone construction (pp. 552-555). Each FOMC press conference audio is manually split into individual answers (692 segments across 36 press conferences). For each audio segment, three audio feature sets are extracted via Librosa:

$$\text{Features}_s = \bigl[\underbrace{128 \text{ mel spectrogram coeff.}}_{\text{loudness at frequency}} ,\; \underbrace{12 \text{ chroma coeff.}}_{\text{melodic/harmonic energy}} ,\; \underbrace{40 \text{ MFCCs}}_{\text{spectral envelope}}\bigr] \quad (180 \text{ features total})$$

A fully connected neural network with architecture: Input(180) -> Linear(200) -> Linear(200) -> Linear(200) -> Softmax(5) [with dropout rate 0.3 after each hidden layer] is trained on RAVDESS and TESS labeled emotion databases (80% train / 20% test). Accuracy on the test set is 84% overall (equation 1, p. 553):

$$\text{Accuracy}(y,\hat{y}) = \frac{1}{n} \sum_{i=1}^{n} \mathbf{1}\{\hat{y}_i = y_i\} \tag{1}$$

with per-class accuracy: angry 87%, sad 84%, neutral 74%, surprised 87%, happy 80%. Each segment is then collapsed to positive (happy or pleasantly surprised), negative (sad or angry), or neutral. Press-conference-level VoiceTone is (equation 2, p. 554):

$$\text{VoiceTone} = \frac{\text{Positive answers} - \text{Negative answers}}{\text{Positive answers} + \text{Negative answers}}, \qquad \text{VoiceTone} \in [-1, +1] \tag{2}$$

The measure is constructed for three Fed chairs: Bernanke (12 press conferences), Yellen (16), and Powell (8). Mean VoiceTone is +0.64 (Bernanke), -0.13 (Yellen), -0.30 (Powell); within-chair variation is substantial (Table 1, p. 558).

Step 2: TextSentiment via BERT (pp. 556-557). BERT base model (12 layers, 768 hidden states, 110M parameters) generates word embeddings for each paragraph of FOMC policy texts. A neural network trained on hand-scored statements from 1997-2010 (scored -10 to +10 on dovish-hawkish axis by research assistants; 81% accuracy) classifies each paragraph as dovish (score $\geq +0.5$), hawkish (score $\leq -0.5$), or neutral. TextSentiment aggregates over statements, opening remarks, and Q&A responses (equation 3, p. 557):

$$\text{TextSentiment} = \frac{\text{Dovish text} - \text{Hawkish text}}{\text{Dovish text} + \text{Hawkish text}}, \qquad \text{TextSentiment} \in [-1, +1] \tag{3}$$

A positive value signals a dovish/expansionary stance. TextSentiment and VoiceTone are weakly correlated ($\rho \approx 0.29$ for Q&A text vs. voice tone; $\rho \approx 0.37$ (Pearson; Spearman = 0.30) for statement text vs. voice tone), confirming partial independence (pp. 557-558; Figure 1, p. 559).

This method builds on text-classification for the BERT-based text scoring and speech-emotion-recognition (proposed) for the audio emotion classification.

Empirical specifications

Main specification (local projections, Jorda 2005). For each horizon $h = 0, 1, \ldots, 15$ days, equation (4) from p. 560 is estimated separately by OLS:

$$\text{Outcome}_{t,t+h} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_t + b_2^{(h)} \text{TextSentiment}_t + b_3^{(h)} \text{FFRShock}_t + b_4^{(h)} \text{FGShock}_t + b_5^{(h)} \text{APShock}_t + b_6^{(h)} \text{ShadowRate}_t + b_7^{(h)} \mathbf{1}\{\text{NoPressConference}_t\} + \text{error}_t^{(h)} \tag{4}$$

where $t$ indexes FOMC meetings; $\text{Outcome}_{t,t+h}$ is the log-return from open on FOMC day $t$ to close on day $t+h$ (e.g. $\log(\text{SPY}_{t+h}^{\text{close}}) - \log(\text{SPY}_t^{\text{open}})$); FFRShock, FGShock, APShock are Swanson (2021) policy shocks (normalized to unit variance over the pre-ZLB period); ShadowRate is from Wu and Xia (2016); and $\mathbf{1}\{\text{NoPressConference}\} = 1$ when no press conference was held. The coefficient path $\{b_1^{(h)}\}_{h=0}^{15}$ traces the impulse response to a unit increase in voice tone. Standard errors are from nonparametric accelerated bootstrap (90% bias-corrected confidence intervals; p. 561). The sample has 68 FOMC meetings, 36 with press conferences.

Results (R1)-(R6) in the Core results table are from specification (4). Outcomes studied: SPY (R1, R6), VIX (R2), VIXY, VIXM (Figure 5, p. 564-565), GOVT (R3), GOVT-TIP spread (R4), EUR/USD (R5), and others (LQD, GLD, IVR, JPY).

Robustness variants (pp. 570-575; Figure 14, p. 574). Adding the Citigroup Economic Surprise Index (Panel A), controlling for volume of corporate earnings announcements (Panel B), adding pre-FOMC media sentiment (Panel C), adding Fed chair fixed effects (Panel D), allowing sign asymmetry in voice tone (Panel E), and replacing BERT with RoBERTa (Panel F), FinBERT (Panel G), search-and-count (Panel H), or human classification (Panel I) leave the voice tone coefficient on SPY stable and statistically significant.

High-frequency (intraday) specifications (pp. 575-577). Audio segments are timed to the second; SPY prices are matched to each answer. “Flow” specification (eq. 6’, p. 575), answer by answer:

$$\text{Outcome}_{t \to t+h,m,s} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_{t,m,s} + b_2^{(h)} \text{TextSentiment}_{t,m,s} + \lambda_m + \gamma_s + \text{error}^{(h)} \tag{6'}$$

“Cumulative” specification (eq. 6”, p. 575), cumulating from press conference start $t_0$:

$$\text{Outcome}_{t_0 \to t+h,m,s} = b_0^{(h)} + b_1^{(h)} \text{VoiceTone}_{t_0 \to t,m,s} + b_2^{(h)} \text{TextSentiment}_{t_0 \to t,m,s} + \lambda_m + \gamma_s + \text{error}^{(h)} \tag{6''}$$

where $\lambda_m$ is a meeting fixed effect and $\gamma_s$ is a question-order fixed effect; $h$ is in minutes; bootstrap clusters at the meeting level. Result R7 uses specification (6’).

Datasets used

Dataset	Role in paper	Wiki page
FOMC press conference audio (Fed YouTube, 36 conferences, April 2011-June 2019)	Construct VoiceTone: 692 Q&A audio segments, split per answer; author-introduced dataset	No page yet
FOMC statements, opening remarks, Q&A transcripts (Fed website)	Construct TextSentiment; training corpus from 1997-2010 statements	No page yet
RAVDESS (Ryerson Audio-Visual Database of Emotional Speech and Song)	Train emotion-detection neural network (80% of combined training set)	No page yet
TESS (Toronto Emotional Speech Set)	Train emotion-detection neural network (joint with RAVDESS)	No page yet
ETF prices: SPY, VIX, VIXY, VIXM, GOVT, TIP, LQD, LQDH, IVR, GLD (Thomson Reuters, Yahoo Finance, Tiingo)	Outcome variables: daily log returns from open on FOMC day to close on day t+h	No page yet
Swanson (2021) policy shocks (FFR, FG, AP)	Controls for monetary policy actions in specification (4)	No page yet
Wu and Xia (2016) shadow policy rate	Control for ZLB period and monetary policy cycle in specification (4)	No page yet
Nexis Uni news database (23,275 FOMC-related articles, Jan 2011-Jul 2019)	Robustness: MediaSentiment control for media framing before each FOMC meeting	No page yet
Federal Reserve Twitter (Board of Governors + regional Fed accounts)	Robustness: post-FOMC tweet sentiment tracks positive voice tone (Figure 16, p. 578-579)	No page yet

Sample: 68 FOMC meetings, 36 press conferences, April 2011 to June 2019. VoiceTone mean = 0.09, SD = 0.75 across all meetings (Table 1, p. 558).

When to read the full paper

Use the original (or preprint) if you are: studying nonverbal central bank communication as a channel for market expectations; building or extending an emotion-detection model for audio from policy events; evaluating whether text-only NLP models fully capture FOMC communication; checking exact robustness specifications and the Internet Appendix with additional results by ETF maturity and currency; or replicating via the ICPSR replication data. The locators above point to the exact figures and tables.

Attribution and rights

Source: peer-reviewed, The American Economic Review 113(2), February 2023. Paywalled; preprint available at SSRN. This distillation was extracted by an LLM on 2026-06-25 and is not human-verified or independently reproduced. Rights are held by the American Economic Association; no CC licence was found in the Crossref record. Reproduce and adapt only per AEA terms.

Gorodnichenko, Yuriy, Tho Pham, and Oleksandr Talavera. “The Voice of Monetary Policy.” The American Economic Review 113, no. 2 (February 2023): 548-584. DOI: 10.1257/aer.20220129.